# VANDAM, the full site for agents > Every page of vandam.ai as Markdown, in one file. Short map: https://vandam.ai/llms.txt # VANDAM · AI transformation studio > AI transformation, from the way you work to the systems you run. New York. URL: https://vandam.ai Contact: hendrik@vibelab.be # Your company. Redesigned. AI transformation, from the way you work to the systems you run. ## About 1 in 20 companies gets real value from AI. Every company can buy the same models. Almost none of them see it in their results. - **5%** of companies get value from AI at scale. Most of the rest report hardly any. [BCG](https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap), 2025 - **33%** Language models could take on 94% of the tasks in computer and math jobs. Claude covers 33% of them. [Anthropic](https://www.anthropic.com/research/labor-market-impacts), March 2026 ## From the first idea to everyday use. Four ways we help. Start with the part your business needs. ### Strategy Choose where AI can help and what to start first. ### Implementation Put existing tools and custom software into daily use. ### Engineering Build with the data, quality checks and connections you need. ### Education Learn, build together and keep up with what changes. ## The org chart is being redrawn. [Anthropic](https://www.anthropic.com/research/employment-at-anthropic) studied its own workforce in October 2026. Agent hours now outnumber human hours **6 to 1**, and around **80%** of an agent session is spent waiting for a person. Human attention has become the bottleneck. Hiring managers talk more about judgment and autonomy, and less about domain expertise. One employee put it this way: “that 20% that was strategic becomes 90% of the job.” [Read our notes on the report](https://vandam.ai/notes/inside-anthropic) - …the intelligence tools we’re creating and using, paired with smaller and flatter teams, are enabling a new way of working which fundamentally changes what it means to build and run a company.: **Jack Dorsey** Co-founder and CEO Block February 2026 - No pure managers: Every leader at: Coinbase must also be a strong and active individual contributor. **Brian Armstrong** Co-founder and CEO Coinbase May 2026 - As engineering, product, design, DS, etc. melt into a new kind of role, I was reflecting on what roles might look like in the future.: **Boris Cherny** Head of Claude Code Claude Code, Anthropic June 2026 - The best way to get real efficiency gains is you have to actually reengineer the process to take advantage of what agents are actually good at.: **Aaron Levie** CEO Box September 2026 ## The motor went where the engine stood. When factories got electricity, most kept the old floor plan and saw little change. The gains came once they were rebuilt around the new power, as [Paul David](https://en.wikipedia.org/wiki/Paul_A._David) of Stanford showed in [The Dynamo and the Computer](https://gwern.net/doc/economics/automation/1990-david.pdf) (1990). AI is at the same point. [Why AI licenses rarely make a company faster](https://vandam.ai/notes/adoption-vs-diffusion) ### 1900s Factories swapped the steam engine for a motor and kept the line shafts. Little changed until each machine got its own motor and the floor followed the work. ### Today Most companies add AI to the same seven steps and make one of them faster. The gains come when the work is redrawn around agents and people decide the exceptions. ### VANDAM We redraw the floor plan first, then put the agents in. ## Years in the business. Young builders make the AI. ### Read We have spent years in finance, operations, marketing and leadership. That is how we pull the business, people and finance logic out of your operations, logistics and back office. ### Redraw With that logic on paper, we redraw the map: which steps go, which merge and where the hours come back. ### Build Then young AI builders make the agents and systems on the new map, and hand them over to your team. [Why we hire HR and finance](https://vandam.ai/notes/ai-needs-hr-and-finance) - Hendrik De Winne: Co-founder · AI and marketing - Thomas De Winne, CFA: Co-founder · Finance and strategy [Meet the team](https://vandam.ai/team) ## Ways to start. ### AI Opportunity Map $9,600 ### Company Brain $19,200 ### Invoice to ERP $24,000 [Explore 21 more services](https://vandam.ai/services) 100 Vandam Street New York ## Why VANDAM. ### Photography credit 100 Vandam · [Kidfly182](https://commons.wikimedia.org/wiki/File:Hudson_Square_April_2022_007.jpg) · [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/). Resized; muted colour. ## Built on the model that fits the work. [How we choose models](https://vandam.ai/models) - **Claude** Anthropic - **ChatGPT** OpenAI - **Gemini** Google - **Copilot** Microsoft - **Grok** xAI - **Open source** Llama, Mistral, DeepSeek, Qwen --- # Services · VANDAM > Twenty-four AI services with published prices, from a plan for your next AI projects to systems built into the work and independent audits. URL: https://vandam.ai/services Contact: hendrik@vibelab.be # Twenty-three ways to start. Start with a plan, or go straight to the system you already know you need. The prices come from past projects and work as indications: the final number can move a little with the size and complexity of your company, and we agree it in writing before we start. Every project starts with a [Floor Plan](https://vandam.ai/floor-plan) of how your company works today, credited when you continue. ## Strategy Decide where AI goes first, and end with a plan you can act on. [Read about strategy](https://vandam.ai/approach/strategy) ### Floor Plan How your company works today: people, roles and workflows, mapped before any AI. **$6,400** two weeks · credited if you continue ### AI Opportunity Map A prioritized plan for your next AI projects. **$9,600** two weeks ### Agent-Ready Audit Assess your data, systems and access. ### Agent Relationship Audit How AI tools find, compare and recommend your business. **$11,200** two weeks ### Pilot Workshop Design one AI pilot with your leadership team in a day. **$4,800** one day ## Implementation Systems brought into the work, connected to your tools and handed over to your team. [Read about implementation](https://vandam.ai/approach/implementation) ### Company Brain Company knowledge, with the sources attached. **$19,200** three weeks ### Invoice to ERP Supplier invoices matched and posted to your ERP. Builds on Company Brain. **$24,000** three weeks ### Agent Onboarding Your data, tools and inboxes connected to every new agent. Builds on Company Brain. **$9,600 + $1,600/mo** two weeks · then monthly ### Account Brains Cited research on twenty target accounts. Builds on Company Brain. **$11,200 + $1,600/mo** ongoing · refreshed monthly ### Agent Ops We run your live agents after go-live. **$3,200/mo** monthly · six-month minimum ### Team Harness Roll out Claude to a whole team, safely. **$12,800** three weeks ### Token Ledger AI spend per person, team and model, next to what it produced. ## Engineering Custom software, tests and independent audits, with verified findings and the fixes attached. [Read about engineering](https://vandam.ai/approach/engineering) ### MCP Retrofit Connect your product to Claude, ChatGPT and other MCP clients. **$12,800** two weeks · plus hosting ### Adversarial App Audit Verified security findings with code fixes. **$4,800** one week ### Local Law 144 Audit Bias audits for hiring tools covered by NYC Local Law 144. **$8,000** annual · repeats every year ### Agent Attack Audit Red-team your agents and their connections. **$6,400** two weeks ### Eval Suite Test every new model on your own work. ### Private AI Run sensitive steps on a model you control. **$16,000** three weeks · plus hardware or hosting at cost ### Source Cleanup Clean the data your agents will read. ## Education Training and advice on your own work, so the skills stay with your people. [Read about education](https://vandam.ai/approach/education) ### 100×worker Program Role assessments and training on your own work. **$8,000** two weeks · up to 50 people, $120 per person beyond ### Judgment Vault Expert decisions, documented and tested. Builds on Company Brain. **$16,000** four weeks · per decision maker ### AI Colleague A weekly call on what changed in AI and what it means for you. **$2,400/mo** weekly call · monthly, stop any month ### Personal AI Setup Your own assistant, set up properly. **$4,800** one week · per person ## All of it, together Take enough of these services and the company itself works differently. A Company Redesign maps how the work runs, picks the services that fit and builds them in order, so its scope and price follow the size of the company. ### Company Redesign Redesign how the whole company works around AI. **From $64,000** ten weeks or more, scoped to the size of the company --- # 100×worker Program by VANDAM: Role assessments and training on your own work > Practical AI training shaped around your team’s tasks. You receive a role-by-role assessment, sessions focused on the gaps it reveals, and a follow-up assessment using the same method to see what changed. URL: https://vandam.ai/100x-worker-program Contact: hendrik@vibelab.be # Your team, using AI on its own work. Practical AI training shaped around your team’s tasks. You receive a role-by-role assessment, sessions focused on the gaps it reveals, and a follow-up assessment using the same method to see what changed. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20100%C3%97worker%20Program) **$8,000** two weeks · up to 50 people, $120 per person beyond ## People already use AI at work. Few were taught how. - **78%** of AI users bring their own AI tools to work, and 52% are reluctant to admit using it for their most important tasks. [Microsoft and LinkedIn](https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part), 31,000 workers, 2024. - **25%** faster, with 12.2% more tasks done, for BCG consultants using GPT-4 on work inside AI’s reach. Outside it, they were 19% less likely to get it right. [Harvard Business School](https://www.hbs.edu/faculty/Pages/item.aspx?num=64700), 2023. - **30s** for ad work that used to take 30 minutes, after a growth marketer at [Anthropic](https://claude.com/blog/how-anthropic-uses-claude-marketing) with no coding background built their own tools. > I’ve seen many of these people approach implausible tasks, ones we wouldn’t even have chosen to tackle before, with reflexive and brilliant usage of AI to get 100X the work done. > > Tobi Lütke, CEO, Shopify, April 2025, [Source](https://www.cnbc.com/2025/04/07/shopify-ceo-prove-ai-cant-do-jobs-before-asking-for-more-headcount.html) So yes, reflexive use can be learned. People practice on their own tasks until it sticks. ## Make the training relevant to the next working day. Buying licences is the easy part. [Vas](https://x.com/vasuman/status/2085806422072418632), who runs Varick Agents, described an operations team of thousands that rolled out Claude Cowork. “The team kept moving at the same rate.” He sees the same barbell in every organization he works with, from 50 people to 5,000. So we start from the roles and the tasks in your company. People practise on their own work: writing useful instructions, checking the output against the source and deciding which calls stay theirs. The same assessment runs again after the sessions, so you can see who moved and where the team still needs help. Power users 5-10% Use it poorly 20% Never use it 70% Employees after an enterprise AI rollout, in the split Vas sees across organizations. Source: [Vas, Varick Agents](https://x.com/vasuman/status/2085806422072418632) , August 2026. ## From the role assessment to practical use. 1. Assess the roles: We map the work in each role and where AI could contribute. The initial assessment creates a comparable starting point across departments and identifies the tasks and knowledge gaps to address. Role map · account manager - Call prepAI can help a lot - ProposalsAI drafts, you edit - Negotiationstays human - CRM updatesAI can help a lot 2. Practise on your work: Sessions use your team’s own tasks and examples. Participants work through an exercise, check the result against its source and decide what to revise or handle themselves before using the output. Exercise · your own proposal - Draft from last quarter’s briefdone - Check every price against the source2 wrong - Rewrite the opening yourselfdone - Sendyour call 3. Repeat the assessment: We reassess using the same method after training. You receive the role maps, team baseline and follow-up view, showing what changed and where further practice would be useful for the team. Team baseline, before after - Finance41 68 - Sales45 74 - Operations38 52 - HR33 61 ## Trained once. Assessed again. Tools change, roles shift and new people join who missed the sessions. After the program we can run the same assessment again and plan practice where the scores slip. - Reassessment: The same assessment repeated, so every round compares with the first baseline. - New joiners: People who join later get the role map and the exercises for their role. - Updated exercises: Exercises rewritten when your tools or tasks change. - Practice sessions: Sessions on the gaps the latest assessment shows. [Ask about ongoing practice](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20100%C3%97worker%20Program%20care) ## What you receive. - Role maps: Task-level assessments for each role. - Team baseline: A comparable view across departments. - Practical training: Sessions designed around the gaps in the assessment. - Follow-up assessment: The same method, repeated after training. $8,000 two weeks · up to 50 people, $120 per person beyond An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For leadership and people teams at 20–200 person companies. Useful when AI adoption varies across roles and departments. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20100%C3%97worker%20Program) ## Where the program leads. Each next step has its own service. ### Team Harness Rolls out the tools to the whole team, safely, with the skills from the program. **$12,800** ### Personal AI Setup A full setup for the people who need more than training. **$4,800** ### Judgment Vault Writes down what your experts know, so juniors learn from it. **$16,000** ### AI Colleague Keeps leadership current every week after the program ends. **$2,400/mo** ## Before you ask. ### Do people need to be technical? No. The exercises use tasks people already do, such as customer emails, proposals and reports. ### How do you measure progress? The same assessment runs before and after the sessions, by role and department. You see the baseline and the follow-up side by side. ### What material do you need from us? Examples of everyday work for each role, agreed with you before the sessions. We use only the examples you approve. ### Does everyone get the same training? No. Sessions follow the gaps in each role’s assessment, so a finance controller and an account manager practice different tasks. --- # Account Brains by VANDAM: Cited research on twenty target accounts > Current, searchable research on twenty target accounts. Your sales team gets company changes, leadership information and buying context with the sources attached, refreshed monthly and available through the AI tools they use. URL: https://vandam.ai/account-brains Contact: hendrik@vibelab.be # Go into the meeting knowing the account. Current, searchable research on twenty target accounts. Your sales team gets company changes, leadership information and buying context with the sources attached, refreshed monthly and available through the AI tools they use. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Account%20Brains) **$11,200 + $1,600/mo** ongoing · refreshed monthly ## Reps spend most of the week on everything except selling. - **40%** of a rep’s week goes to selling. The other 60% goes elsewhere. [Salesforce](https://www.salesforce.com/en-us/wp-content/uploads/sites/4/documents/reports/sales/salesforce-state-of-sales-report-2026.pdf) survey of 4,050 sales professionals, 2026. - **16%** of their time goes to preparation and planning, by the same [Salesforce](https://www.salesforce.com/en-us/wp-content/uploads/sites/4/documents/reports/sales/salesforce-state-of-sales-report-2026.pdf) survey. - **59%** of B2B buyers feel reps don’t take the time to understand their business. [Salesforce](https://www.salesforce.com/news/stories/sales-ai-statistics-2024/), 2024. > I’m able to cover all these accounts thanks to a skill that runs as a scheduled task overnight. > > John Albert, Business development, Anthropic, August 2026, [Source](https://claude.com/blog/how-anthropics-business-development-team-uses-claude-to-run-inbound-and-outbound-at-scale) So no, nobody has to research twenty accounts by hand before every call. ## Stop rebuilding the research before every call. Before a meeting, a rep opens the company website, scans the news and guesses who signs off. Next week a colleague does the same research on the same account. What they found ends up in browser tabs, private notes and a slide that is out of date by next quarter. We keep that research in one place per account, with the source next to every fact. A rep can ask what changed since the last call, open the announcement it came from and walk in with a better first question. Where the public record runs out, the brief says so and turns the gap into a question for the meeting. Account Brains build on your [company brain](https://vandam.ai/company-brain): each account page links to your own people, offer and past decisions. ## How the account brief stays useful. 1. Research the target accounts: We compile leadership information, company announcements, product changes and published plans for the twenty accounts. Findings are linked to their sources so a rep can check the original context. Calder & Finch · sources - Second plant announced, ColumbusNewsroom, 3 Oct - Three new roles at the new siteCareers page, 9 Oct - New COO from grocery distributionLeadership page, 21 Oct - Packaging cost named a priorityAnnual report, p. 14 2. Make the research accessible: We give each rep access through a connected AI tool and organize the findings around the account. A half-day training session uses your own accounts to practise preparing for calls. Ask about an account What changed at Calder & Finch since my last call? They announced a second plant and hired a COO from grocery distribution. Packaging suppliers get chosen before the new line runs. - Newsroom, 3 Oct - Annual report, p. 14 3. Refresh and keep the history: Monthly updates recheck the sources and record what changed. The team keeps a shared account history instead of losing the previous research whenever someone prepares a new meeting brief. Monthly refresh · November - 20 accounts rechecked412 sources - Calder & Finch2 changes - Northcote FixturesCFO left - 17 accountsno change ## Researched once. Rechecked every month. Accounts change between calls: people move, plans get announced and pages disappear. The monthly fee keeps every account current and records what changed, so a rep always works from this month’s research. - Source checks: Every source on the twenty accounts is rechecked each month. - Change record: What changed per account since last month, dated and linked to its source. - Open questions: New gaps in the public record are added to the brief as questions for the next call. - Access: We run the research and each rep’s MCP connection to their AI tool. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Account%20Brains%20care) ## What you receive. - Twenty account profiles: Connected research with checkable sources. - Individual access: An MCP connection for your reps’ AI tools. - Team training: A half-day session using your own accounts. - Monthly updates: Fresh source checks and a record of what changed. The same engine runs newyorktech.com, a cited graph of the New York tech scene. [View related work](https://newyorktech.com) $11,200 + $1,600/mo ongoing · refreshed monthly An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For B2B sales teams at 20–100 person companies. Useful when account research is repeated before every meeting. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Account%20Brains) ## Where Account Brains leads. Each next step has its own service. ### Company Brain Comes before: your own people, offer and decisions the account pages link to. **$19,200** ### Agent Onboarding Sales agents load the accounts they work on, with the right access. **$9,600 + $1,600/mo** ### Agent Relationship Audit How AI tools describe you to those same accounts when they compare suppliers. **$11,200** ## Before you ask. ### Which accounts are included? Twenty target accounts that you choose with us at the start. ### Where does the research come from? Public sources such as company announcements, leadership and careers pages, published plans and reports. Every finding links to its source, and where the public record runs out the brief says so. ### Which AI tools can our reps use? AI clients that support MCP connections. Each rep gets individual access, and the half-day training uses your own accounts. ### What if a fact is wrong or out of date? Every fact shows its source and date, so a rep can check it before using it. The monthly refresh rechecks each source and records what changed or disappeared. --- # Adversarial App Audit by VANDAM: Verified security findings with code fixes > A focused security review of your application. For each verified finding, you receive steps to reproduce the issue, the affected code and a proposed fix, together with the checks behind the report. URL: https://vandam.ai/adversarial-app-audit Contact: hendrik@vibelab.be # Verified security findings. Code fixes included. A focused security review of your application. For each verified finding, you receive steps to reproduce the issue, the affected code and a proposed fix, together with the checks behind the report. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Adversarial%20App%20Audit) **$4,800** one week ## One missing check is the most common flaw on the web. - **#1** Broken access control tops the [OWASP Top 10](https://top10.owasp.org/2025/A01_2025-Broken_Access_Control), the security industry’s standard list of web app risks, 2025. - **#1** Reading another customer’s record by changing an ID tops the [OWASP API Security Top 10](https://api-security.owasp.org/editions/2023/en/0xa1-broken-object-level-authorization), rated easy to exploit and widespread. - **$10.22M** average cost of a data breach in the US, a record. [IBM](https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls), Cost of a Data Breach, 2025. > Our research shows models are getting better at coding accurately but are not improving at security. > > Jens Wessling, CTO, Veracode, July 2025, [Source](https://www.veracode.com/press-release/ai-generated-code-poses-major-security-risks-in-nearly-half-of-all-development-tasks-veracode-research-reveals/) So no, code that passes its tests can still show one customer another’s data. ## Give engineering something it can act on. A security report that says access control is weak leaves engineering with the real work still to do: find where it fails and decide what to change. A useful finding names the request that breaks it, the code responsible and the fix. We test sign-in, data access, storage and payments within an agreed scope, the places where one missing check exposes another customer’s records. Every issue is reproduced, then rechecked before it reaches the report. Your engineers get the reproduction, the affected files, a proposed code change and a test that shows the hole is closed. ## From a suspected issue to a reviewed fix. 1. Agree and map the scope: We identify the application boundaries to review, including who can sign in, which records they can access and how storage and payments are handled. Testing stays within the agreed environment and scope. Scope - Sign-in · email + Googlein scope - Customer recordsin scope - File storagein scope - Payments · Striperead only 2. Reproduce the finding: We test realistic misuse cases and recheck any suspected issue. A finding includes the request or action that demonstrates it and the relevant files, so another engineer can follow the evidence. Finding 03 · reproduced - GET /api/invoices/8812 - signed in as: customer A - 200 OK · returns customer B’s invoice - file: api/invoices/[id].ts, line 14 3. Deliver the code change: We prepare replacement code with the location and reason for the change. The review record keeps the original reproduction, supporting evidence and checks together so your team can assess the fix. Fix · api/invoices/[id].ts - const invoice = await db.invoice.find(id) - const invoice = await db.invoice.find({ id, - ownerId: session.user.id }) - if (!invoice) return notFound() ## Tested this week. Retested after each release. The report describes your application as it was during the audit week. Every release can add a route or remove a check, so after delivery we can rerun the reproductions and test what changed. - Fix verification: We rerun the original reproduction against each fix your team ships, and record the result in the finding. - Release retests: Before or after a release, we retest the findings and routes it touches. - New scope: New roles, routes, storage or payment flows reviewed with the same method as the first audit. - Check-ins: A short session with engineering on open findings, accepted risks and what the next release changes. [Ask about retesting](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Adversarial%20App%20Audit%20care) ## What you receive. - Verified findings: Reproduced issues linked to the relevant files and lines. - Code fixes: Replacement code for the findings we verify. - Review record: The scope, evidence and checks behind the report. $4,800 one week An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For founders and engineering teams with a live application. Useful before enterprise review or after a security concern. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Adversarial%20App%20Audit) ## Where the audit leads. Each next step has its own service. ### Agent Attack Audit When the app has agents or AI features: the same adversarial testing, aimed at them. **$6,400** ### Agent Ops Keeps watch after the fixes, so the same findings do not come back. **$3,200/mo** ### MCP Retrofit Once the app holds up: an interface assistants and agents can use safely. **$12,800** Developers with an AI assistant wrote significantly less secure code, and were more likely to believe it was secure. ([Stanford](https://arxiv.org/abs/2211.03622) study by Perry, Srivastava, Kumar and Boneh, 2023) ## Before you ask. ### Do you test on our production system? We agree the environment and the test accounts before we start, and testing stays inside them. Anything outside the agreed scope is left alone. ### What if you find nothing serious? The report says so. The review record shows what we tested and how, so the result can be checked. ### Do our engineers apply the fixes? Yes. We deliver replacement code with the location and reason for each change, and your team reviews and merges it. ### Can we share the report in a customer’s security review? The report is yours to share. It covers the scope we agreed and says so, which matters if the customer asks about other parts of your system. --- # Agent Attack Audit by VANDAM: Red-team your agents and their connections > A findings report with a reproduction for every attack that worked, and the fixes already applied and re-tested. We test the support bots, email agents, Cowork connectors and MCP servers you run today, with the access they hold today. URL: https://vandam.ai/agent-attack-audit Contact: hendrik@vibelab.be # See what your agents do when a stranger gives the orders. A findings report with a reproduction for every attack that worked, and the fixes already applied and re-tested. We test the support bots, email agents, Cowork connectors and MCP servers you run today, with the access they hold today. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent%20Attack%20Audit) **$6,400** two weeks ## Agents read text from strangers. Few companies limit what they can reach. - **#1** risk in the [OWASP](https://genai.owasp.org/llmrisk/llm01-prompt-injection/) Top 10 for LLM applications, 2025: prompt injection, including instructions hidden in websites and files a model reads. - **97%** of organizations that reported a breach of an AI model or application had no AI access controls in place. [IBM](https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls) Cost of a Data Breach Report, 2025. - **10+ days** of data exports from Salesforce instances in August 2025, using stolen tokens from the Salesloft Drift chat integration. [Google Threat Intelligence](https://cloud.google.com/blog/topics/threat-intelligence/data-theft-salesforce-instances-via-salesloft-drift). > Prompt injection is the new #1 attack vector. > > Alex Lieberman, Co-founder, Tenex, June 2026, [Source](https://x.com/businessbarista/status/2069529729309512135) ## An agent reads what strangers send. Then it acts with your keys. A support bot reads tickets written by customers. A payables agent reads invoices written by suppliers. A language model has no reliable way to tell your instructions apart from instructions hidden in that text, and [OWASP](https://genai.owasp.org/llmrisk/llm01-prompt-injection/) ranks this, prompt injection, first in its Top 10 for LLM applications. Its advice starts with limiting what the model can reach to what the task needs. That reach decides the damage. In August 2025, attackers used stolen tokens from the Salesloft Drift chat integration to export data from many companies’ Salesforce instances, then searched it for AWS keys and Snowflake tokens, according to [Google Threat Intelligence](https://cloud.google.com/blog/topics/threat-intelligence/data-theft-salesforce-instances-via-salesloft-drift). A chat tool with wide access to the CRM became the way into it. We test the agents you already run the way an attacker would: planted instructions in tickets, emails, shared documents and tool descriptions, then we watch which tools the agent reaches for and what its tokens allow. What gets through is fixed in the same two weeks and attacked again. It is the agent counterpart of the [Adversarial App Audit](https://vandam.ai/adversarial-app-audit), which does the same for the code of an app. > Every agent with access to your systems is a new door into your company, and a lot of those doors were left open. > > Greg Isenberg, CEO, Late Checkout, October 2026, [Source](https://x.com/gregisenberg/status/2110414696037716372) ## How the two weeks run. 1. Map what each agent can reach: We list every agent, connector and MCP server, the content each one reads and the scopes and keys it holds. The first finding is often a token with far more access than the job needs. 2. Attack with your own kind of content: We write attacks in the shapes your agents meet every day: a support ticket, an invoice email, a shared document, the description of a third-party tool. Each attempt is logged with the agent’s reasoning and every tool call it made. 3. Fix, then attack again: Scopes are cut to the task, outside content reaches the model as data, keys move out of the agent’s reach and risky actions wait for a person. Every attack that worked runs again until it fails. ## Attacked once. Attacked again after every change. Models update, connectors get added and prompts change, and any of those can reopen a door. After the audit we can run the attack set every month and test each new agent before it goes live. - Attack rerun: The full attack set run against every agent, next to last month’s results. - New connections: Every new agent, connector or MCP server attacked before it goes live. - Scope review: Tokens and permissions compared with what each agent actually used. - Findings call: A monthly session with IT on what changed and what to close next. [Ask about monthly attack runs](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent%20Attack%20Audit%20care) ## What you receive. - Attack map: Every agent and connection with the access it really has. - Findings report: Each finding with the input, the tool call and a severity. - Fixes, re-tested: Changes applied and the same attacks run again until they fail. $6,400 two weeks An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For operations, IT and security leads at 20–500 person companies with agents in production. Useful once an agent reads outside content and can act in your systems. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent%20Attack%20Audit) ## Where the audit leads. Each next step has its own service. ### Agent Onboarding Reconnects each agent with only the access it needs, owner by owner. **$9,600 + $1,600/mo** ### Eval Suite Turns every attack that worked into a test case run on each new model. **$12,800** ### Agent Ops We run the agents with the fixes in place and watch for the next attempt. **$3,200/mo** ### Private AI Moves the most sensitive steps to a model you control. **$16,000** > a permissions manager for your AI agents. which agents can access which accounts, what’s the spending limit, what requires your approval. nobody is building this and everyone is going to need it. > > Greg Isenberg, CEO, Late Checkout, April 2026, [Source](https://x.com/gregisenberg/status/2048381278953603406) ## Before you ask. ### Do you attack our production systems? We agree the scope with your IT lead first. Most tests run against the live agents with test accounts and test records, so an attack that works shows what would have happened without touching real customers or money. ### Which agents and tools do you cover? Support bots, email and payables agents, Claude Cowork connectors, MCP servers, custom GPTs and in-house agents: anything that reads outside content or holds a key. We start from the inventory and test what can reach the most. ### How does this differ from the Adversarial App Audit? The Adversarial App Audit tests the code of an app you built. This audit tests agents: what they read, how they can be steered and what their scopes let them reach. ### Can prompt injection be fixed for good? No model is immune today. The fixes limit what a steered agent can do: narrow scopes, outside content passed as data, keys out of reach and a person approving risky actions. The re-test shows what still gets through. --- # Agent Onboarding by VANDAM: Your data, tools and inboxes connected to every new agent > We connect your documents, APIs, MCP servers, Gmail and WhatsApp numbers once, and give each agent a role, a manager and limits. A new agent starts work the day you need it. URL: https://vandam.ai/agent-onboarding Contact: hendrik@vibelab.be # Every system connected once. Every new agent ready on day one. We connect your documents, APIs, MCP servers, Gmail and WhatsApp numbers once, and give each agent a role, a manager and limits. A new agent starts work the day you need it. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent%20Onboarding) **$9,600 + $1,600/mo** two weeks · then monthly ## Your people already brought their own agents. - **78%** of AI users bring their own AI tools to work, per the 2024 [Work Trend Index](https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part). - **58%** of workplace Claude use runs through personal accounts, per the data-security vendor [Cyberhaven](https://www.cyberhaven.com/blog/sensitive-data-flowing-into-ai-tools), February 2026. - **13/13** frontier models were hijacked at least once in a public red-teaming competition run with [NIST’s CAISI](https://www.nist.gov/blogs/caisi-research-blog/insights-ai-agent-security-large-scale-red-teaming-competition), March 2026. ## Every new agent starts with the same week of setup. An agent is only as useful as what it can reach. The order desk needs the shared inbox, the WhatsApp number, the shop and the ERP; the finance agent needs invoices, the bank feed and the price lists. Without a plan, each agent gets its own pasted API keys and someone’s personal login, and after the third one nobody can say what reaches what. We do the setup once, as a layer every agent uses. Each source is connected with its own scopes, and each agent gets an identity file: its role, the person it reports to, the knowledge it may load, its working hours and the gates on anything that sends a message or spends money. Our notes on [what agents can reach](https://vandam.ai/notes/what-your-agents-can-reach) and [internal controls](https://vandam.ai/notes/agents-and-internal-controls) explain why those limits matter. Agent Onboarding builds on your [company brain](https://vandam.ai/company-brain): each agent loads its own slice of it, by role, on day one. ## From scattered keys to agents with a desk. 1. Map and connect: We list every source, API, MCP server, inbox and number with its owner, then connect each one once, with read, draft and send scopes kept apart. Connection map - Gmail · orders@read, draft - WhatsApp Business · +32 9read - Shopify · MCPread - ERP APIread - Send to customersoff 2. Give every agent an identity: Each agent gets a role, a manager, a knowledge slice, working hours and gates. Messages to customers and anything that spends money start switched off until a person approves them. Iris · order desk agent - **Manager** Mara, customer service - **Knows** Orders, stock, delivery terms - **Works** Mon to Fri, 7:00 to 19:00 - **Can** Read orders, draft replies - **Needs approval** Sending, refunds, discounts 3. Keep it current every month: New agents are onboarded on the same layer, keys are rotated, and a monthly access review records what each agent reached and what changed. Access review · October - 2 new agents onboardedFinance, Sales - Keys rotated14 of 14 - Unused access removedDrive · Iris - Written review sentto IT lead ## Set up in two weeks. Kept current every month. Agents get added, people change roles and keys need rotating, so the layer needs regular upkeep. The monthly fee covers that work. - New agents: New agents onboarded on the same layer, each with a role card, a manager and gates. - Key rotation: Keys and tokens rotated every month, with the date recorded on each connection. - Access review: A written review of what each agent reached, what changed and which unused access was removed. - Check-ins: A monthly session with your IT lead on the review, open requests and the next agents to onboard. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent%20Onboarding%20care) ## What you receive. - Connection map: Every data source, API, MCP server, inbox and number, with what can reach it. - Agent identities: A role, a manager, a knowledge slice and gates for each agent, in files you own. - First agents live: Up to three agents onboarded and tested on real work. - Monthly upkeep: New agents onboarded, keys rotated and a written access review every month. $9,600 + $1,600/mo two weeks · then monthly An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For operations and IT leads at 30–500 person companies starting to run agents. Useful when every new agent means another week of keys, connectors and access questions. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent%20Onboarding) ## Where Agent Onboarding leads. Each next step has its own service. ### Company Brain Comes before: the company knowledge each agent loads on day one. **$19,200** ### Agent Attack Audit Red-teams the connections you just opened, before someone else does. **$6,400** ### Agent Ops We run the agents after go-live, with the access map kept current. **$3,200/mo** ### Token Ledger Spend per agent, team and model, next to what each one produced. **$9,600 + $1,600/mo** The only way to stay safe there is to avoid that lethal trifecta combination entirely. (Simon Willison, [June 2025](https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/)) ## Before you ask. ### Can an agent send a customer message or spend money on its own? Not by default. Sending and spending start switched off, and a person approves them before they are turned on. Each agent’s gates are written in its identity file. ### What happens to the keys and logins people already use? We list them in the mapping step with their owners. Each source is then connected once with its own scopes, so agents no longer run on someone’s personal login. ### How many agents are included? Up to three agents are onboarded and tested on real work in the first two weeks. New agents after that are onboarded as part of the monthly upkeep. ### What does the monthly access review contain? Which agents reached which sources, which keys were rotated and what access was removed or changed. It is written down and sent to your IT lead. --- # Agent Ops by VANDAM: We run your live agents after go-live > A monthly retainer to run your live agents. We read their traces, turn every failure into a fix and a test, re-run the eval suite whenever a model or your data changes, and send a KPI report each month and a shortlist of what to automate ne URL: https://vandam.ai/agent-ops Contact: hendrik@vibelab.be # Keep your agents working long after go-live. A monthly retainer to run your live agents. We read their traces, turn every failure into a fix and a test, re-run the eval suite whenever a model or your data changes, and send a KPI report each month and a shortlist of what to automate next each quarter. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent%20Ops) **$3,200/mo** monthly · six-month minimum ## Most teams run agents in production. About half test them. - **57%** of organizations have agents running in production. [LangChain](https://www.langchain.com/state-of-agent-engineering), State of Agent Engineering, survey of 1,340 professionals. - **52%** run offline evaluations on test sets, and 37% run online evals on real-world performance. [LangChain](https://www.langchain.com/state-of-agent-engineering), same survey. - **32%** name quality as a top barrier to production, the most cited barrier in the survey. [LangChain](https://www.langchain.com/state-of-agent-engineering). > Turn on tracing first. Traces = receipts of every action (tool calls, Salesforce pings, web searches, dead ends). > > Alex Lieberman, Co-founder, Tenex, September 2026, [Source](https://x.com/businessbarista/status/2099565601312166157) The trace shows where it went wrong, usually a few steps before the error. ## Agents drift after launch. Someone has to own them. An agent that worked in its first week meets new cases every week after. A supplier changes its invoice layout, a model version is retired, a field in the CRM gets renamed. The agent keeps running while some of its results quietly get worse, and without someone reading the traces, the first sign is a customer or a controller finding the mistake. Agent Ops is that someone. We review the traces of every run, the record of each step and tool call, and catch failures, loops and slow drift. Each failure becomes a fix and a new case in the eval suite, so the same mistake cannot come back unnoticed. When a model or a data source changes, the whole suite runs again before the change reaches your team. “Who owns a build after it’s deployed?” is on the list of questions [Alex Lieberman](https://x.com/businessbarista/status/2077143133049995619)’s team hears most from enterprise executives. Many companies will hire for it. [Aaron Levie](https://x.com/levie/status/2043883641366032638) describes the role as an agent deployer and manager who “manages evals and reviews after any major model or data change, and runs and manages the agents on an ongoing basis tracking KPIs.” Until that person is in place, you can rent the job by the month, and hand it over when they start. > The more enterprises I talk to about AI agent transformation, the more it’s clear that there is going to be a new type of role in most enterprises going forward. The job is to be the agent deployer and manager in teams. > > Aaron Levie, CEO, Box, April 2026, [Source](https://x.com/levie/status/2043883641366032638) ## How a month of Agent Ops runs. 1. Read the traces: Tracing is switched on for every agent we run, in your own observability tool or in one we set up in your environment. We review runs every working day, flag failures, loops and cost spikes, and keep a log of what we found. 2. Fix and test: Each failure is traced to its cause: a prompt, a tool, a data change or a model change. We write the fix, add the failing case to the eval suite and re-run the suite before the fix goes live through your release process. 3. Report and look ahead: At the end of each month you receive the KPI report per agent: runs, completion rate, failures, time to fix and hours saved. Each quarter we bring a ranked shortlist of what to automate next, with the expected gain. ## Run by us. Yours to take in-house. The six-month minimum gives the eval suite and the runbook time to cover the cases your agents actually meet. Everything lives in your environment: traces, evals, fixes and the runbook. When you hire your own agent manager, we hand it over and train them. - Trace review: Runs read every working day, failures logged with their trace. - Eval re-runs: The full suite run after every model, prompt or data change. - KPI report: Monthly, per agent, next to the month before. - Handover: Runbook, evals and training for the person who takes over in-house. [Ask about Agent Ops](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent%20Ops%20care) ## What you receive. - Trace review: Every agent run read for failures, loops and quiet errors. - Failure-to-fix loop: Each failure traced, fixed and added to the eval suite. - Eval re-runs: The full suite re-run after any model, prompt or data change. - KPI report: Runs, completion rate, time to fix and hours saved, per agent. $3,200/mo monthly · six-month minimum An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For operations and IT leads with agents in production. Useful once an agent runs every day and nobody has time to watch it. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent%20Ops) ## Where Agent Ops leads. Each next step has its own service. ### Eval Suite Tests a new model on your own cases before it reaches the live agents. **$12,800** ### Token Ledger What every live agent costs, next to what it produced. **$9,600 + $1,600/mo** ### Agent Attack Audit A yearly red-team of the agents we run and their connections. **$6,400** > Right now nobody knows which of their agents are actually good and which ones are quietly wasting money. > > Greg Isenberg, CEO, Late Checkout, April 2026, [Source](https://x.com/gregisenberg/status/2048381278953603406) ## Before you ask. ### Which agents can you run? Agents we built and agents built by your team or another partner, as long as tracing can be switched on and the agent can run against an eval suite. If there is no suite yet, building one is part of the first month. ### Do you need access to production? Read access to traces and logs, and a test environment for the evals. Fixes go live through your own release process, with your sign-off. ### What counts as a failure? A run that ended in an error, produced a wrong result, looped or cost far more than usual. We agree the thresholds per agent in the first month. ### What happens when a model is updated? The eval suite runs on the new version before your agents switch. You see the result next to the current version and decide. --- # Agent-Ready Audit by VANDAM: Assess your data, systems and access > A readiness report, a scorecard and a prioritized list of fixes. You learn which data, process, connection and permission gaps stand between your company and a useful first AI project, with the evidence attached. URL: https://vandam.ai/agent-ready-audit Contact: hendrik@vibelab.be # Know what needs fixing before the AI build. A readiness report, a scorecard and a prioritized list of fixes. You learn which data, process, connection and permission gaps stand between your company and a useful first AI project, with the evidence attached. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent-Ready%20Audit) **$9,600** two weeks ## Most companies can’t say whether their data is ready. - **63%** of organizations lack, or don’t know if they have, the data practices AI needs. [Gartner](https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk) survey of 1,203 data leaders, 2025. - **30/50** AI practitioners interviewed by [RAND](https://www.rand.org/pubs/research_reports/RRA2680-1.html) described persistent data quality problems, the second most common reason projects fail. - **23%** of organizations are scaling AI agents anywhere, while 62% are experimenting. [McKinsey](https://www.mckinsey.com/no/our-insights/the-state-of-ai), State of AI, November 2025. > Every company I talked to wants to pipe its internal data into one place and expose it to agents through MCP. Then they look at the data and realize it’s a mess: five versions of the same customer, old Slack channels, PDFs nobody can search. > > Greg Isenberg, CEO, Late Checkout, October 2026, [Source](https://x.com/gregisenberg/status/2110414696037716372) So no, the mess is normal. We find out which part blocks the first agent. ## Find the dependencies before they delay the project. An AI agent is software that carries out tasks in your connected tools. Before it does useful work it needs information it can read, a process it can follow and permission to take the right actions. A demo can look convincing with one of the three missing, and the gap tends to surface halfway through the build. We trace the workflow you have in mind back to everything it depends on. The report separates the foundations you can build on from the gaps, explains what each fix makes possible and names who has to be involved. You start the build knowing its real scope. > Data has to be cleaned up and moved to modern systems, processes have to be re-engineered and documented, workflows have to be evaled, SLAs have to get achieved, and so on. > > Aaron Levie, CEO, Box, June 2026, [Source](https://x.com/levie/status/2067455756795039957) ## How the readiness report is assembled. 1. Inspect the current setup: We inventory your sources, systems, process instructions and access arrangements. Representative records and conversations with the team reveal how the work happens and where the proposed system would need help. Inventory - CRM · 18,400 recordssampled 200 - Order inboxread 3 weeks - Process notes2 of 6 written down - Admin accessone person holds it 2. Assess each dependency: We assess readiness across data, process, tools and permissions. Findings link back to the evidence reviewed, so your technical and operating teams can understand the reason behind each part of the score. Readiness by area - Data71 - Process44 - Tools82 - Permissions39 3. Put the fixes in order: We rank the actions by the work they enable, their expected effort and their dependencies. You receive the score breakdown, evidence file and an ordered list of improvements to discuss with owners. Fixes, in order - Write the returns process downunblocks 3 builds - Split the shared admin logintwo days - Clean duplicate CRM accountsone week - Add API access to the ERPvendor quote ## Scored once. Re-scored every quarter. Fixes land, systems change and the next workflow comes up. After the report we can re-run the assessment each quarter, so the score shows what moved and what still stands between you and the build. - Re-score: Each area assessed again with the same checks, next to last quarter’s score. - Fix tracking: Every fix followed from open to closed, with its owner and the date it closed. - Next workflow: A new workflow scored against what its agent would need, before anyone scopes the build. - Review session: A quarterly session with the owners on what moved, what stalled and what to fix next. [Ask about quarterly re-checks](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent-Ready%20Audit%20care) ## What you receive. - Readiness assessment: A 0–100 score with a breakdown by category. - Evidence file: Findings linked to the sources and systems reviewed. - Prioritized fixes: Actions with estimated effort and dependencies. $9,600 two weeks An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For leadership and operations teams at 20–200 person companies. Useful before connecting agents to internal systems. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent-Ready%20Audit) ## Where the audit leads. Each next step has its own service. ### Floor Plan Comes before the audit: the workflows it checks your data and systems against. **$6,400** ### Source Cleanup Fixes the data the audit flagged, so agents read one clean version of it. **$12,800** ### Agent Onboarding Connects the systems that passed, with an owner and the right access for every agent. **$9,600 + $1,600/mo** ### MCP Retrofit When your own product is the gap: an interface assistants and agents can use. ## Before you ask. ### Which workflow is the score measured against? One you choose with us at the start, usually the first task you want an agent to take on. Every area is scored against what that agent would need. ### What access do you need to our systems? An inventory, representative records and a view of how access is set up. We agree with your IT lead in advance what we look at and how. ### Does the audit include fixing the gaps? No. You receive the score, the evidence file and an ordered list of fixes with owners and effort. The build can be scoped once the gaps are known. ### What if our score is low? Then you know before the build budget is spent. The fix list separates the gaps that block the first workflow from the ones that can wait. --- # Agent Relationship Audit by VANDAM: How AI tools find, compare and recommend your business > A report on how AI tools find, describe and compare your business, plus recordings of buying tasks on your site. You receive the observed gaps and a prioritized list of changes your team or agency can make. URL: https://vandam.ai/agent-relationship-audit Contact: hendrik@vibelab.be # See what buyers hear when they ask AI. A report on how AI tools find, describe and compare your business, plus recordings of buying tasks on your site. You receive the observed gaps and a prioritized list of changes your team or agency can make. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent%20Relationship%20Audit) **$11,200** two weeks ## Buyers ask an assistant before they ask you. - **69%** of software buyers chose a different vendor than planned because a chatbot recommended it. [G2](https://company.g2.com/news/g2-research-the-answer-economy) survey of 1,076 B2B buyers, April 2026. - **51%** now start their research with an AI chatbot more often than with Google, by the same [G2](https://company.g2.com/news/g2-research-the-answer-economy) survey. - **1%** of visits to a Google results page with an AI summary ended in a click on a link inside the summary. [Pew Research Center](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/), 68,879 searches by 900 US adults, 2025. > AI chatbots have compressed weeks of vendor research into a single prompt. > > Tim Sanders, Chief Innovation Officer, G2, April 2026, [Source](https://company.g2.com/news/g2-research-the-answer-economy) So no, you don’t get those weeks back. We read what that one prompt says. ## Understand the route before the buyer contacts you. A buyer can ask an AI assistant for a shortlist of suppliers before ever visiting your site. The answer may lean on an old product page, leave out your delivery area or repeat a competitor’s claim. By the time the buyer calls, they have already met that version of your company. We put the questions that matter in your category to four AI models with search switched on, and keep what they answer and which pages they cite. Then we have an assistant attempt an agreed task on your site and on three competitors’ sites, and record where it gets stuck. Every finding comes with the citation or recording behind it and the change that would fix it, so your marketing team gets a list of edits it can make. ## How we examine the buyer’s experience. 1. Ask the buying questions: We run thirty relevant buying questions across four models with search enabled. We retain the answers and cited sources so your team can inspect how your company and category were represented. Buying question · model 2 of 4 Which distributor can deliver packaging to Brooklyn by Friday? Northline Supply and Harbor Pack both list next-day delivery in New York. Atlas & Co. is not mentioned. - northline.com/delivery - Review site, 2024 2. Test the next step: We observe agreed tasks on your site and three competitors’ sites. Recordings show the steps reached, missing information and blockers, including whether a form or request gives a useful confirmation. Quote request · atlas.example - Found the quote form2 clicks - Product sizes listedmissing - Form submittedok - Confirmation with next stepnone sent 3. Prioritize the improvements: We combine the answer evidence and task observations into an assessment and action plan. Recommendations identify what to change, why it matters and which recorded finding supports the work. Assessment - Reach34 - Readability58 - Evidence41 - Task completion62 ## Measured once. Asked again every quarter. Models change, competitors update their pages and your own changes go live. Running the same questions and tasks again each quarter shows which answers moved and which still need work. - Questions re-run: The same thirty questions put to the four models again, with each answer next to last quarter’s. - Tasks re-tested: The agreed tasks recorded again on your site and the three competitors’ sites. - Change tracking: Each change from the action plan followed against the answers it was meant to correct. - Review session: A quarterly session with Marketing on what moved, what did not and what to change next. [Ask about quarterly re-checks](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent%20Relationship%20Audit%20care) ## What you receive. - Visibility assessment: A 0–100 assessment of reach, readability, evidence and task completion. - Session recordings: Your site and three competitors tested on the same tasks. - Shortlist map: Thirty buying queries across four models, with cited sources. - Action plan: Prioritized changes your team or agency can implement. $11,200 two weeks An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For commercial and marketing teams at 50–500 person companies. Useful when buyers research your category through AI tools. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Agent%20Relationship%20Audit) ## Where the audit leads. Each next step has its own service. ### MCP Retrofit Lets assistants act on your product, not only mention it: look up, compare and order. **$12,800** ### Company Brain Keeps the facts AI tools should repeat about you in one place, with the sources attached. **$19,200** ### AI Colleague Follows how the assistants change week to week, and what that means for how they find you. **$2,400/mo** ## Before you ask. ### Can you get us recommended? Nobody can promise what a model will say. The audit shows what the models say today, which pages they rely on and what you can change on your side. ### Who chooses the questions and tasks? We draft the buying questions for your category and agree the tasks with you before anything is run. ### Do you make the changes? The action plan is written for your team or agency to implement. Each item names the page, the change and the finding behind it. ### Why test competitors’ sites? A blocker means more when you can see whether a competitor’s site gets the same assistant through the same task. All four sites are tested on the same tasks. --- # AI Colleague by VANDAM: A weekly call on what changed in AI and what it means for you > A weekly call of 30 to 60 minutes, with a one-page brief sent the day before. We read the model launches, the updates to tools you already pay for, the new rules and what your industry is doing, check all of it against what your company run URL: https://vandam.ai/ai-colleague Contact: hendrik@vibelab.be # Hear what changed in AI each week and what it means for your company. A weekly call of 30 to 60 minutes, with a one-page brief sent the day before. We read the model launches, the updates to tools you already pay for, the new rules and what your industry is doing, check all of it against what your company runs and is building, and tell you what to try, what to watch and what to ignore. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20AI%20Colleague) **$2,400/mo** weekly call · monthly, stop any month ## AI ships faster than companies can read it. - **88%** of organizations now use AI. [Stanford HAI](https://hai.stanford.edu/ai-index/2026-ai-index-report), AI Index 2026. - **1 year** for the best models to go from 60% to near 100% on SWE-bench Verified, a standard coding test. [Stanford HAI](https://hai.stanford.edu/ai-index/2026-ai-index-report), same report. > At dinner, someone asked the table, “If a new model drops next month, does your company get better or worse?” HALF the table went quiet! > > Greg Isenberg, CEO, Late Checkout, October 2026, [Source](https://x.com/gregisenberg/status/2110414696037716372) The headline someone forwarded on Monday often ends up under ignore. ## More ships each week than anyone can read. One person reads it with your company in mind. Every week a lab ships a model, Microsoft and Google add features to tools you already pay for, ChatGPT and Claude change what they can connect to, and a regulator publishes guidance. Most of it has nothing to do with your company. The few items that change what your team should try this month arrive in the same stream as everything else. Without a filter the stream reaches the company anyway, through whoever reads the most. A board member forwards a headline, a competitor announces a strategy, and a decision gets made on what was in the news that week instead of on what fits your systems. We do the reading. Each week Hendrik or a named VANDAM lead goes through the releases with your profile next to them: the tools you run, your industry, the projects you have open and the questions your team sent in. You get a one-page brief and a call. When something on the call is worth building, it becomes a scoped project, often a Pilot Workshop or an Agent-Ready Audit. If a general read is enough for now, VANDAM Weekly is the free version: one email a week that Hendrik writes for every company at once. [Read VANDAM Weekly](https://vandam.ai/weekly) ## How a week runs. 1. Read the week: From Monday to Friday we follow model launches, updates to Microsoft 365, Google Workspace, ChatGPT, Claude and the other tools you use, open-source models, regulation such as the EU AI Act, and what companies in your industry announce. 2. Check it against your company: Each item is held against your profile: your stack, your industry, your open projects and the questions your team sent in. What passes gets a line on what it means for you and a mark: try, watch or ignore. Most items do not pass. 3. Brief, then call: The one-page brief arrives the day before. On the call we go through it, answer what came up during the week and agree who tries what by when. The decisions go into the brief archive. ## A record that builds up. Yours to keep. Every brief, call note and decision is stored in your own workspace, so the record stays with the company when people change roles. The profile is updated as your stack and projects change. You can stop any month and keep everything. - Weekly call: 30 to 60 minutes with Hendrik or a named VANDAM lead. - One-page brief: Sent the day before, every item marked try, watch or ignore. - Questions in between: By message, whenever something comes up during the week. - Profile updates: Your stack, projects and questions updated as they change. [Ask about AI Colleague](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20AI%20Colleague%20care) ## What you receive. - Weekly call: 30 to 60 minutes with Hendrik or a named VANDAM lead. - One-page brief: What changed, what it means for you, and what to try, watch or ignore. - Questions in between: Ask by message when something lands in the middle of the week. - Brief archive: Every brief and decision kept, a record of what you chose and why. $2,400/mo weekly call · monthly, stop any month An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For CEOs, leadership teams and the people who own AI at 20–500 person companies. Useful when AI news moves faster than anyone in the company can read it, and decisions start to follow the headlines. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20AI%20Colleague) ## Where the weekly call leads. Each next step has its own service. ### AI Opportunity Map When the calls keep returning to the same ideas: rank them and pick the first. **$9,600** ### Pilot Workshop Turns one idea from the calls into a pilot in a day. **$4,800** ### Team Harness Brings what you learn on the calls to the whole team. **$12,800** ## Before you ask. ### Who is on the call? Hendrik or a named VANDAM lead on our side. On yours, usually the CEO and whoever owns AI, sometimes the full leadership team. It works best with the same people each week. ### What do you need from us to filter? A first conversation about the tools you run, your industry, the projects you have open and the questions on the table. From there we keep the profile current with you. ### Is there a free version? Yes. VANDAM Weekly is a general weekly email that Hendrik writes for every company at once. AI Colleague is the same reading, checked against your company, with a call. ### Will you push us toward a vendor? We work model-agnostic and on the stack you already have. VANDAM is a Claude implementation partner and often suggests Claude first, and we say so when that is the case. Most weeks the advice is about tools you already pay for. ### What happens when something is worth building? We scope it as a separate project with its own price, usually a Pilot Workshop or an Agent-Ready Audit first. The call stays the place where you decide what to start. ### Can we stop? Any month. The brief archive stays with you. --- # AI Opportunity Map by VANDAM: A prioritized plan for your next AI projects > A ranked plan for AI projects across your departments: what to build first, how to build it, and which tools you already have do the job. A standalone report your own team, or your coding agent, can build from. URL: https://vandam.ai/ai-opportunity-map Contact: hendrik@vibelab.be # Know which AI projects are worth starting. A ranked plan for AI projects across your departments: what to build first, how to build it, and which tools you already have do the job. A standalone report your own team, or your coding agent, can build from. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20AI%20Opportunity%20Map) **$9,600** two weeks ## Most AI pilots stop before they reach the business. - **46%** of AI proofs of concept were scrapped before production, on average. 42% of companies abandoned most of their AI initiatives in 2025, up from 17%. [S&P Global](https://www.ciodive.com/news/ai-project-failure-rates/742590/), 1,000+ companies. - **50%+** of generative AI budgets go to sales and marketing tools, while the biggest returns showed up in back-office automation. [MIT NANDA](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/), 2025. > pick one pain point, execute well, and partner smartly with companies who use their tools. > > Aditya Challapally, Lead author, MIT, August 2025, [Source](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/) So no, more pilots won’t fix it. Picking the right first one does. ## Turn the ideas into a decision. Most leadership teams have a list of AI ideas and no good way to rank it. The loudest request often solves a small problem, and a convincing demo can hide that the data behind it does not exist yet. Pick the wrong first project and the company spends a year learning that AI does not work here. We rank every candidate on the same terms: the hours it takes off someone’s week, what has to be ready first, and what it costs to build and run. The founder of Varick Agents, quoted by [Annelies Gamble](https://x.com/AnneliesGamble/status/2069814714214023186), a partner at Zetta Venture Partners, puts the cut-off bluntly: “A workflow with a 10x ROI should get prioritized. A workflow with a 2x ROI maybe, maybe not. And a workflow with a sub 1x ROI should be left alone.” We write down why each idea starts, waits or stops, so the people approving the budget and the people doing the work argue from the same page. - **5%** of companies get value from AI at scale. [BCG](https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap), September 2025. - **60%** report hardly any material value from it. Same study. - **3×** High performers are nearly three times as likely to have redesigned their workflows. [McKinsey](https://www.mckinsey.com/capabilities/operations/our-insights/the-state-of-ai), 2025. ## How the priorities are agreed. 1. Start with the departments: We interview the people doing and managing the work, using consistent questions. They show us recurring tasks, repeated effort and the situations that regularly need attention. Interviews - Finance · 4 peopleinvoice coding, 9 h/week - Sales · 3 peoplequote prep, 6 h/week - Customer service · 5order questions, 14 h/week - Operations · 2scheduled Thursday 2. Compare the opportunities: We assess potential impact, preparation and implementation effort. Each candidate is considered alongside its data, connections and process requirements so the comparison includes the work a demonstration can hide. Impact vs. effort - Order questionshigh impact · low effort - Invoice codinghigh · medium - Quote prepmedium · medium - Demand forecastlow · data missing 3. Agree the starting point: We review the recommendations with your team. You receive ranked opportunities, a view by department and the reasoning behind the order, ready to discuss a defined first build. Recommendation - Start: order questions14 h/week back - Next: invoice codingafter ERP access - Wait: quote prepprocess unclear - Leave: demand forecastno clean data ## Ranked once. Re-ranked each quarter. The map is yours to keep. Data gets cleaned, systems change and new ideas arrive, so we can come back each quarter and update the ranking on the same terms. - Re-ranking: Every opportunity scored again against what changed since the last review. - New ideas: Ideas raised during the quarter assessed on the same terms as the original list. - Started projects: The builds that started, compared with the hours and effort in the plan. - Leadership session: A quarterly session to agree what starts next, what waits and what to drop. [Ask about quarterly reviews](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20AI%20Opportunity%20Map%20care) ## What you receive. - Priority map: Ranked opportunities with the reasoning attached. - Impact and effort: An assessment for each opportunity. - Build specs: What to build and how for every project to start, or the tool that already does it. $9,600 two weeks An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For leadership teams at 20–250 person companies. Useful when there are several ideas and no agreed starting point. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20AI%20Opportunity%20Map) ## Where the map leads. Each next step has its own service. ### Floor Plan Comes before the map: how your company works today, so the opportunities are ranked on the real workflows. Its fee is credited. **$6,400** ### Agent-Ready Audit For the opportunities marked to start: is your data, are your systems and access ready to carry them, and what to fix first. **$9,600** ### Company Brain Most maps put company knowledge near the top. The brain is the base the next systems build on. **$19,200** ### Pilot Workshop Takes the first opportunity and designs it as a pilot with your leadership team in a day. **$4,800** ### Company Redesign When the map shows the change runs across the company: the services that fit, built in order. **From $64,000** > AI isn’t delivering ROI yet, but we’re all in, so we need to figure it out. > > Garrett Lord, CEO, Handshake, June 2026, [Source](https://x.com/GarrettLord/status/2068754262440767500) ## Before you ask. ### Do we need you to build the projects? No. The map ends with a build spec for each project to start, written so your own developers, or a coding agent like Claude Code, can work from it, and it names the existing tools where nothing needs building. If you want us to build, the first build is scoped as a separate project. ### What do we have at the end? Ranked opportunities with the reasoning attached, an impact and effort assessment for each, a build spec or tool pick for every opportunity marked to start, and a view by department. ### How much of our team’s time does it take? Interviews with the people who do and manage the work in each department, and a review session with leadership at the end. We agree the interview list with you before we start. ### What if our favorite idea ranks low? The plan says why: too few hours saved, too much to prepare first, or too much to run. Ideas move up when their prerequisites are in place, and a quarterly review picks that up. --- # Approach · VANDAM > Strategy, implementation, engineering and education: the four ways VANDAM works with a company, each with its own services. URL: https://vandam.ai/approach Contact: hendrik@vibelab.be # One company, four ways in. Strategy picks the work to change. Implementation brings it into your team, engineering builds what your tools cannot do, and education leaves the skills behind. Each one can start on its own. ## [Strategy](https://vandam.ai/approach/strategy) Decide where to invest in AI and keep those decisions current. We assess the opportunities, build the roadmap and work with leadership as projects develop and priorities change. [Floor Plan](https://vandam.ai/floor-plan) [AI Opportunity Map](https://vandam.ai/ai-opportunity-map) [Agent-Ready Audit](https://vandam.ai/agent-ready-audit) [Agent Relationship Audit](https://vandam.ai/agent-relationship-audit) [Pilot Workshop](https://vandam.ai/pilot-workshop) [Read about strategy](https://vandam.ai/approach/strategy) ## [Implementation](https://vandam.ai/approach/implementation) Introduce the right combination of existing tools and custom software into your business. We coordinate the setup, connections, process changes and coaching that help your team use it day to day. [Company Brain](https://vandam.ai/company-brain) [Invoice to ERP](https://vandam.ai/invoice-to-erp) [Agent Onboarding](https://vandam.ai/agent-onboarding) [Account Brains](https://vandam.ai/account-brains) [Agent Ops](https://vandam.ai/agent-ops) [Team Harness](https://vandam.ai/team-harness) [Token Ledger](https://vandam.ai/token-ledger) [Read about implementation](https://vandam.ai/approach/implementation) ## [Engineering](https://vandam.ai/approach/engineering) Build custom applications, automations and AI agents with the data, testing and operating controls they need. We develop the software, check its behavior and prepare your team to run it. [MCP Retrofit](https://vandam.ai/mcp-retrofit) [Adversarial App Audit](https://vandam.ai/adversarial-app-audit) [Local Law 144 Audit](https://vandam.ai/local-law-144-audit) [Agent Attack Audit](https://vandam.ai/agent-attack-audit) [Eval Suite](https://vandam.ai/eval-suite) [Private AI](https://vandam.ai/private-ai) [Source Cleanup](https://vandam.ai/source-cleanup) [Read about engineering](https://vandam.ai/approach/engineering) ## [Education](https://vandam.ai/approach/education) Develop your team's AI skills through practical training, building together and continuing advice. We work on the tasks that matter to your people, with support that fits their experience and ambitions. [100×worker team assessment](https://vandam.ai/100x-worker-program) [Judgment Vault](https://vandam.ai/judgment-vault) [AI Colleague](https://vandam.ai/ai-colleague) [Personal AI Setup](https://vandam.ai/personal-ai-setup) [Read about education](https://vandam.ai/approach/education) --- # Education · VANDAM > Develop your team's AI skills through practical training, building together and continuing advice. We work on the tasks that matter to your people, with support that fits their experience and ambitions. URL: https://vandam.ai/approach/education Contact: hendrik@vibelab.be [Strategy](https://vandam.ai/approach/strategy) [Implementation](https://vandam.ai/approach/implementation) [Engineering](https://vandam.ai/approach/engineering) [Education](https://vandam.ai/approach/education) # Give your team the skills to use AI. Develop your team's AI skills through practical training, building together and continuing advice. We work on the tasks that matter to your people, with support that fits their experience and ambitions. [Plan learning for your team](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Education) ## Practical skills and continued support. - **Role-based learning** Training matched to people's tasks, current skills and the tools they use. - **Reusable working material** Instructions, review guidance and the work developed together, available for the team to keep using. - **Internal confidence** Support for colleagues who will guide others, introduce new starters and help the team continue learning. ## Judgment is becoming the skill that counts. [Anthropic](https://www.anthropic.com/research/employment-at-anthropic) looked at how its own hiring managers assess candidates. For technical roles, the biggest increases were in comments on **autonomy and judgment**, while domain expertise shrank as a share of the assessment. By 2026, mentions of creative AI use had **tripled** for technical roles and **quadrupled** for non-technical roles. [Boris Cherny](https://x.com/bcherny/status/2071379474277613732), who leads Claude Code at Anthropic, sees five kinds of people on his team: the prototyper, the builder, the sweeper, the grower and the maintainer. Many people span two of them, sometimes three. I also notice that these roles are not really tied to job function. ([Boris Cherny](https://x.com/bcherny/status/2071379474277613732), head of Claude Code at Anthropic, June 2026) So we train judgment, whatever the job title. People learn to choose suitable tasks, give clear instructions, check the result and know which information they may use, on the work their own role is responsible for. Experience keeps developing after a session, so you choose the rhythm: focused learning, building together or an ongoing AI colleague. ## How your team learns. 1. Choose the focus: You share the team's roles, tools and priorities. We assess the starting point and agree the learning format, the work to bring and who should participate. 2. Work alongside your people: Participants practice, build and review with us. We explain decisions as we go, make room for questions and help people apply what they learn to their own responsibilities. 3. Decide what comes next: Together we review progress and the questions that remain. Further coaching, building sessions or advisory calls can be agreed around the support your team wants. ## Choose how we work together. Start with a focused session, build a project together or work with us regularly. ### 100×worker program Weekly sessions for individuals and teams who want to develop their AI skills over time. We map your work, bring together the information your tools need and build on real tasks. Each session develops your ability to use, adapt and improve the result independently. [Explore 100×worker](https://100xworker.com/) ### Workshops and ongoing training Focused sessions or continued training on agreed tools, tasks and ways of working. Participants use suitable material from their own work, compare results and leave with guidance they can apply again. ### Build a project with us For teams with a defined project, we build the workflow, automation or tool together. Your people participate in the decisions, development and testing, learning how to maintain and extend the result as it takes shape. ### An ongoing AI colleague Weekly or monthly calls combine relevant AI updates with advice on your current work. We explain what new capabilities mean for your team, review questions and obstacles, and help decide what to try or learn next. ## Learn with the tools you use Training follows your work, whether you are writing, building or running a process. - **[Wispr Flow](https://wisprflow.ai/)** Voice dictation - **[Lovable](https://lovable.dev/)** Build web apps - **[Codex](https://openai.com/codex/)** Build and review code - **[Claude](https://claude.com/)** Write, research and analyze ## Services for this step ### 100×worker team assessment A fixed assessment, training and reassessment package for your team. **$8,000** two weeks · up to 50 people, $120 per person beyond ### Judgment Vault Expert decisions, documented and tested **$16,000** four weeks · per decision maker ### AI Colleague A weekly call on what changed in AI and what it means for you **$2,400/mo** weekly call · monthly, stop any month ### Personal AI Setup Your own assistant, set up properly **$4,800** one week · per person **Start with a conversation** One email to Hendrik, then a price for the service that fits, agreed before we start. --- # Engineering · VANDAM > Build custom applications, automations and AI agents with the data, testing and operating controls they need. We develop the software, check its behavior and prepare your team to run it. URL: https://vandam.ai/approach/engineering Contact: hendrik@vibelab.be [Strategy](https://vandam.ai/approach/strategy) [Implementation](https://vandam.ai/approach/implementation) [Engineering](https://vandam.ai/approach/engineering) [Education](https://vandam.ai/approach/education) # Software built around your business. Build custom applications, automations and AI agents with the data, testing and operating controls they need. We develop the software, check its behavior and prepare your team to run it. [Talk about what you need to build](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Engineering) ## Software you can operate and improve. - **Working software** The interface, logic and system connections needed for the agreed task. - **A quality record** Data preparation decisions, repeatable behavior checks and the results used to assess release readiness. - **An operating handover** Code, configuration and documentation, with access, hosting, maintenance and support responsibilities agreed. ## We build for the process you want to run. At [Anthropic](https://www.anthropic.com/research/employment-at-anthropic), agents worked **six times** as many hours as their human handlers by August 2026. An agent is an AI assistant allowed to use connected tools to carry out tasks. That much output only helps a business when the software around the model fits the work: your information, the connections to your systems, the access rules and a clear point where a person checks the result. > The best way to get real efficiency gains is you have to actually reengineer the process to take advantage of what agents are actually good at and can do differently and faster than people. > > Aaron Levie, CEO, Box, September 2026, [Source](https://x.com/levie/status/2096279756925026733) So we agree the process first, then build for it. We define acceptable behavior with your team and check it throughout development: the information the system relies on, the actions it may take and what it does when it cannot finish a task reliably. ## How we build it. 1. Define what good means: Your team brings the requirements, source systems and records of actual work. We agree the expected results, permitted actions and acceptance criteria before building. 2. Build and review: We develop working versions and review them with your team. You try the task, inspect the output and help resolve the business decisions that emerge during development. 3. Release and maintain: We review the checks before release and prepare the operating handover. Ongoing maintenance has an agreed scope, including who responds to failures and approves changes. ## How we check quality. The checks become part of the system, so quality can be reassessed when its data, instructions or software change. ### Data preparation and access We identify the authoritative sources, organize the agreed data and check for missing, conflicting or outdated records. We define how information is retrieved and refreshed, while preserving who is allowed to see it. Data owners help resolve ambiguities before those ambiguities reach users. ### Evaluations An evaluation is a repeatable check of whether the system behaves as intended. We assemble a test set from actual tasks, with expected results or review criteria agreed by your team. It covers output accuracy, supporting sources, permissions and incomplete inputs. Disagreements help identify what needs changing. ### Quality after release For systems we maintain, we rerun the checks when models, instructions, data or code change. This catches updates that weaken earlier results. We monitor failures, response times and usage costs. When your team operates the system, we hand over these checks with the recovery procedures. ## Services for this step ### MCP Retrofit Connect your product to Claude, ChatGPT and other MCP clients **$12,800** two weeks · plus hosting ### Adversarial App Audit Verified security findings with code fixes **$4,800** one week ### Local Law 144 Audit Bias audits for hiring tools covered by NYC Local Law 144 **$8,000** annual · repeats every year ### Agent Attack Audit Red-team your agents and their connections **$6,400** two weeks ### Eval Suite Test every new model on your own work **$12,800** three weeks ### Private AI Run sensitive steps on a model you control **$16,000** three weeks · plus hardware or hosting at cost ### Source Cleanup Clean the data your agents will read **Start with a conversation** One email to Hendrik, then a price for the service that fits, agreed before we start. --- # Implementation · VANDAM > Introduce the right combination of existing tools and custom software into your business. We coordinate the setup, connections, process changes and coaching that help your team use it day to day. URL: https://vandam.ai/approach/implementation Contact: hendrik@vibelab.be [Strategy](https://vandam.ai/approach/strategy) [Implementation](https://vandam.ai/approach/implementation) [Engineering](https://vandam.ai/approach/engineering) [Education](https://vandam.ai/approach/education) # Put AI to work in your team. Introduce the right combination of existing tools and custom software into your business. We coordinate the setup, connections, process changes and coaching that help your team use it day to day. [Talk about your team’s workflow](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Implementation) ## Tools, process and adoption. - **A configured setup** Selected tools and connections, with agreed access, source material and task instructions. - **An operational workflow** A documented way to complete the task, with clear responsibilities, review points and exception handling. - **A supported rollout** Training, workflow coaching and a way to track usage, task outcomes and where people need help. ## Most of the gap sits between the tool and the team. [Anthropic](https://www.anthropic.com/research/labor-market-impacts) measured it in March 2026. Language models could take on **94%** of the tasks in computer and math jobs. In practice, Claude covered **33%** of them. Could do 94% Actually used 33% Share of tasks in computer and math jobs. Source: [Anthropic](https://www.anthropic.com/research/labor-market-impacts) , March 2026. Inside its own company, [Anthropic](https://www.anthropic.com/research/employment-at-anthropic) found where the time goes. Around **80%** of an agent session is spent idle, waiting for human input. The report calls human attention “the bottleneck for agent output.” That waiting is what a rollout has to fix. We decide with the team who supplies the information, who checks the result and who takes over when the system gets stuck. We start from how people work today: the tools they already use, their confidence with AI and the obstacles in their way. ## How the rollout happens. 1. Assess with the team: Employee surveys and conversations with team leads establish the starting point. Your people show us their current tasks, while we review requirements with the relevant system owners. 2. Introduce it to a small group: We configure the agreed tools and coordinate any custom development. A first group uses the process, gives feedback and helps us refine the setup and guidance before wider introduction. 3. Support and adjust: We coach the wider team, review progress against the starting point and address the obstacles that remain. An internal owner takes responsibility for the process, with continuing support agreed around your needs. ## Choosing what to use. We assess the task, the systems it touches and the people who will run it before settling on the tooling. ### Existing tools We assess capabilities, permissions, data handling, integration options and recurring costs. Configuration includes the material the tool can use, the instructions it follows and how it fits the team's existing accounts and responsibilities. ### Custom software When existing products cannot meet the requirements, we scope and build custom software. Implementation covers how it reaches users, connects to their tools and becomes part of their daily work. ### A combination that can be maintained Existing products and custom components can work together. We make their boundaries and dependencies clear, identify who maintains each part and account for the effect of updates. Coaching and user feedback continue to inform the setup. ## Tools we work with We choose around the task, your team and the systems you already use. - **[Wispr Flow](https://wisprflow.ai/)** Voice dictation - **[Lovable](https://lovable.dev/)** Build web apps - **[Codex](https://openai.com/codex/)** Build and review code - **[Claude](https://claude.com/)** Write, research and analyze - **[Clay](https://www.clay.com/)** Research and enrich accounts - **[ChatGPT](https://chatgpt.com/overview/)** Research, writing and analysis - **[n8n](https://n8n.io/)** Connect tools and automate work - **[Cursor](https://cursor.com/)** Develop software with AI ## Services for this step ### Company Brain Company knowledge, with the sources attached **$19,200** three weeks ### Invoice to ERP Supplier invoices matched and posted to your ERP **$24,000** three weeks ### Agent Onboarding Your data, tools and inboxes connected to every new agent **$9,600 + $1,600/mo** two weeks · then monthly ### Account Brains Cited research on twenty target accounts **$11,200 + $1,600/mo** ongoing · refreshed monthly ### Agent Ops We run your live agents after go-live **$3,200/mo** monthly · six-month minimum ### Team Harness Roll out Claude to a whole team, safely **$12,800** three weeks ### Token Ledger AI spend per person, team and model, next to what it produced **Start with a conversation** One email to Hendrik, then a price for the service that fits, agreed before we start. --- # Strategy · VANDAM > Decide where to invest in AI and keep those decisions current. We assess the opportunities, build the roadmap and work with leadership as projects develop and priorities change. URL: https://vandam.ai/approach/strategy Contact: hendrik@vibelab.be [Strategy](https://vandam.ai/approach/strategy) [Implementation](https://vandam.ai/approach/implementation) [Engineering](https://vandam.ai/approach/engineering) [Education](https://vandam.ai/approach/education) # Set your AI priorities. Decide where to invest in AI and keep those decisions current. We assess the opportunities, build the roadmap and work with leadership as projects develop and priorities change. [Talk about your priorities](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Strategy) ## A clear basis for investment. - **Opportunity assessment** A prioritized view of the work AI could improve, with the business need and readiness of each area. - **Cost and value** Estimates of benefit, effort and ongoing cost, with the assumptions visible and the uncertainties identified. - **A working roadmap** An agreed sequence of work, project owners and decision points that can be updated as you learn. ## Most companies use AI. Few have picked the work to change. In [McKinsey’s 2025 survey](https://www.mckinsey.com/capabilities/operations/our-insights/the-state-of-ai), **88%** of organizations used AI. Far fewer see it in their results. [BCG](https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap) counts only a small group that gets value at scale, and a much larger one that gets almost nothing back. - **5%** of companies get value from AI at scale. [BCG](https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap), 2025. - **60%** reap hardly any material value, with minimal revenue and cost gains. [BCG](https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap), 2025. - **3×** Top performers were nearly three times as likely to have redesigned their workflows. [McKinsey](https://www.mckinsey.com/capabilities/operations/our-insights/the-state-of-ai), 2025. The difference is in which work gets redesigned, and in what order. The [European Central Bank](https://www.ecb.europa.eu/press/key/date/2026/html/ecb.sp260323_1~1e06784a89.en.html) made the same point about whole economies in March 2026: the productivity gain from AI is “primarily driven by **the speed of diffusion**.” Strategy decides where that adoption starts. We weigh business goals, available people and what your systems can support in one discussion. Once projects run, the evidence gets better: what people use, where quality holds and what the systems cost to maintain. We bring that back into the next investment decision. ## How we set the priorities. 1. Set the direction: Leadership shares its goals and constraints. We speak with the people doing the work, review the current systems and agree what the assessment needs to establish. 2. Make the first decisions: Together we compare the opportunities, challenge the estimates and choose the work to pursue. The roadmap records what each project needs before it can begin. 3. Review as the work develops: For an ongoing engagement, we agree a review rhythm with your leadership team. You bring current results and business decisions; we update recommendations and help resolve the next priorities. ## The decisions between projects. Ongoing strategy creates a place to resolve the questions that cross teams, budgets and individual projects, with the right people involved. ### Keep the reasoning We record why a project has priority, what must be true for it to succeed and who makes the next decision. That record makes it possible to revisit a choice when the underlying assumptions change. ### Review the evidence We compare results with the original goals, including output quality, use by the team, operating cost and maintenance effort. The review informs whether to expand the work, change its scope or stop investing. ### Reassess the options We help leadership assess relevant changes in AI capabilities and suppliers against the roadmap. New possibilities receive the same scrutiny as existing projects: business value, readiness, effort and a responsible owner. ## Services for this step ### Floor Plan How your company works today: people, roles and workflows, mapped before any AI **$6,400** two weeks · credited if you continue ### AI Opportunity Map A prioritized plan for your next AI projects **$9,600** two weeks ### Agent-Ready Audit Assess your data, systems and access ### Agent Relationship Audit How AI tools find, compare and recommend your business **$11,200** two weeks ### Pilot Workshop Design one AI pilot with your leadership team in a day **$4,800** one day **Start with a conversation** One email to Hendrik, then a price for the service that fits, agreed before we start. --- # Company Brain by VANDAM: Company knowledge, with the sources attached > A company knowledge assistant that answers everyday questions using the documents, emails and decisions you choose. Your team asks in plain language and can check the original sources. URL: https://vandam.ai/company-brain Contact: hendrik@vibelab.be # Ask a question. Get your company’s answer. A company knowledge assistant that answers everyday questions using the documents, emails and decisions you choose. Your team asks in plain language and can check the original sources. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Company%20Brain) **$19,200** three weeks ## A quarter of the week goes to finding answers. - **25%** of their time, teams say, goes to searching for answers. [Atlassian](https://www.atlassian.com/blog/state-of-teams-2025) survey of 12,000 knowledge workers, 2025. - **62%** struggle with too much time spent searching for information in their workday. [Microsoft](https://www.microsoft.com/en-us/worklab/work-trend-index/will-ai-fix-work) survey of 31,000 workers, 2023. - **95%** of business analytics questions at [Anthropic](https://claude.com/blog/how-anthropic-enables-self-service-data-analytics-with-claude) are now handled by Claude, its data team reports, June 2026. > memory is half the equation. a company brain needs reliable enforcement of the decisions it has made. > > Kasey Klimes, Builds decision infrastructure, ex-Google, Primitive, August 2026, [Source](https://x.com/kaseyklimes/status/2093127970571317258) So no, just dumping all your documents in Claude isn’t a scalable long-term solution. ## The answer exists. Finding it takes the time. A new colleague asks who owns an account. A manager needs to know why a price changed. Someone searches for the current returns policy. The answer may be spread across documents, inboxes and the person who remembers. Company Brain gives your team one place to ask. It brings together the material you choose, answers in everyday language and points to the records behind the answer. That helps people check the detail and understand the context. The brain is also where the next systems start. [Agent Onboarding](https://vandam.ai/agent-onboarding), [Account Brains](https://vandam.ai/account-brains), [Judgment Vault](https://vandam.ai/judgment-vault) and [Invoice to ERP](https://vandam.ai/invoice-to-erp) all build on it. ## How we build it, with a check at every step. Six steps. Each one ends with something you approve before the next starts. 1. Scope and your questions: We agree what goes in, what stays out and the budget, and collect 20 to 30 real questions from at least three roles. Those questions are the test at the end. Test questions - Who can approve a customer discount?Sales lead - How do we handle a damaged return?Support - Why did we change Northwind’s pricing?Finance - Which supplier covers the Ghent site?Operations - 24 more questions from 5 roles 2. Gather and sort: Documents, mailboxes, wikis and tickets are stored as they are and sorted before anything is compiled, so nobody pays to process the wrong material. Sources sorted - Shared drive4,210 files kept - Mailboxes3 of 5 agreed - Wiki312 pages kept - Ticketslast 12 months only 3. Build the org spine: We draft departments, roles and people. You check the org chart before anything is built on top of it. Org spine, draft 2 Atlas & Co. - **Sales** Account team Pricing desk - **Operations** Ghent site Logistics - **Finance** Payables Controlling Waiting for your check 4. Compile the pages: Pages and links are written from your material, and every gap comes back to you as a question. Automated checks must pass before we move on. Compile checks - Every page cites a source1,042 pages - No broken links0 found - Pages without an owner3 sent to you - Stale review dates0 found 5. Wire systems and access: Numbers stay in the systems that own them, each department gets its slice, and anything that sends or spends starts switched off. Access by department - Sales reads Sales + Productsread - Finance figures stay in the ERPlinked - Send emailoff until approved - Spend moneyoff until approved 6. Pass your questions, then hand over: Every question from step one is answered with a citation or recorded as not covered. Your team is trained to add material and keep it current. Ask Atlas Who can approve a customer discount above 10%? The pricing desk lead, up to 15%. Above that, the head of sales signs off in writing. - Pricing policy, v4 - Sales handbook, p. 12 ## Handed over. Still looked after. You own the code and every page from day one. Most teams still want someone to run it, so after handover we can stay on for a monthly fee. - Hosting: We run the brain, its connections to your AI tools and the access rules, on your infrastructure or ours. - Cleanup: A monthly check flags stale pages and broken links, and every correction is swept into a ruling. - Co-building: New sources, departments and questions are added with your team, a few each month. - Check-ins: A monthly session on what people asked and what it missed. When someone leaves, their knowledge is captured while they can still explain it. **$1,600/mo** after handover · monthly [Ask about monthly care](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Company%20Brain%20care) ## What you receive. - Cited knowledge base: Every page links back to its source documents. - Connected tools: Answers through your selected AI tools, subject to your access rules. - Update system: The code, documentation and training to maintain it. $19,200 three weeks An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For leadership and operations teams at 30–300 person companies. Useful for onboarding, recurring questions and knowledge that depends on a few people. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Company%20Brain) ## Where the brain leads. The next systems build on it. ### Source Cleanup Comes before, when the sources disagree: one clean version for the brain to read. **$12,800** ### Agent Onboarding Every new agent loads its own slice of the brain, by role, on day one. **$9,600 + $1,600/mo** ### Judgment Vault Your experts’ decisions become rulings in the brain, next to the policies they apply. **$16,000** ### Account Brains Account pages that link to your own people, offer and past decisions. **$11,200 + $1,600/mo** ### Invoice to ERP Checks every invoice against the agreements and price lists kept in the brain. **$24,000** ## Before you ask. ### Does every company need a graph? No. If the real question is what you decided and how you do things, a handful of well-kept documents can be enough, and we check that before we scope anything. ### Can everyone see every document? Not by design. In a shared folder, roles tell people and agents what is theirs but do not lock anything, so sensitive material stays out; when it has to be in, we deploy with individual logins. ### Can it be wrong? Yes. Every answer cites its pages so a person can check, automated checks catch stale and contradictory pages, and every error we find becomes a ruling so it happens only once. ### Does this replace our existing tools? No. The brain connects to the AI tools you choose. We prefer Claude and also work with other models; your source systems stay where they are. --- # Company Redesign by VANDAM: Redesign how the whole company works around AI > Around ten weeks for a mid-sized company, longer for a larger one, with people who have run finance, operations, marketing and leadership, and young AI builders. We read how your company really works, redraw the map around agents, build the URL: https://vandam.ai/company-redesign Contact: hendrik@vibelab.be # Redraw how the company works and build the systems for it. Around ten weeks for a mid-sized company, longer for a larger one, with people who have run finance, operations, marketing and leadership, and young AI builders. We read how your company really works, redraw the map around agents, build the systems on it and hand them to your people. You end with fewer steps, the hours back in the work and a team that owns what was built. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Company%20Redesign) **From $64,000** ten weeks or more, scoped to the size of the company ## Most of the value comes from redrawing the work. Few companies have done it yet. - **70%** of the value of an AI transformation comes from redesigning workflows, culture, governance and how people and AI work together. The AI application itself accounts for 10%, according to [BCG](https://www.bcg.com/capabilities/artificial-intelligence). - **6%** of companies attribute at least 5% of their EBIT to AI, about the same share as a year earlier. [McKinsey](https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20state%20of%20ai/the-state-of-ai-in-2026-on-the-road-to-roi.pdf), State of AI 2026, 1,719 respondents. - **63%** of finance organizations said AI implementation was slower than expected in 2025. [Gartner](https://cpapracticeadvisor.com/2026/05/28/gartner-says-cfos-must-stop-mistaking-finance-ai-deployment-for-value-creation/184094) survey of 204 finance leaders, March 2026. > The best way to get real efficiency gains is you have to actually reengineer the process to take advantage of what agents are actually good at and can do differently and faster than people. > > Aaron Levie, CEO, Box, September 2026, [Source](https://x.com/levie/status/2096279756925026733) We start with the process everyone complains about. ## Most companies add AI to the steps they already have. The hours come back when the steps change. Most companies we meet have bought AI licenses, run a few pilots and kept the process as it was. Each step gets a little faster, and the work still passes through the same hands, the same approvals and the spreadsheets someone built years ago to bridge two systems. An order still crosses the building as many times as before. Redrawing that takes people who know the business from the inside. We have spent years in finance, operations, marketing and leadership, so we can read the business, people and finance logic out of your operations, logistics and back office, and tell a control the auditors will ask about from a workaround nobody needs anymore. With that logic on paper we decide which steps go, which merge, where an agent takes over and which roles change. Then young AI builders make the agents and systems on the new map, on the stack you already run, and hand them over with the code, the tests and the training. It combines the services on this site that fit your company, and covers what our [approach](https://vandam.ai/approach) splits into strategy, implementation, engineering and education. > Human touch and judgement gets reserved for the first and final mile of most processes. > > Alex Lieberman, Co-founder, Tenex, August 2026, [Source](https://x.com/businessbarista/status/2094213970215231831) ## How the ten weeks run. 1. Read the company: In weeks one to three we sit with the teams in operations, logistics, finance, sales and HR, read the records and the systems, and map every recurring process step by step with the hours it takes. The map shows how the work actually runs, workarounds included. 2. Redraw the map: In weeks three to five every step gets a verdict: it goes, it merges with another, an agent takes it, or a person keeps it because it needs judgment. Leadership signs off on the new map, the hours it returns and the roles that change before anything is built. 3. Build and hand over: In weeks five to ten our builders make the agents and systems in order of hours returned, test them on your own records and put them into the work one team at a time. Every agent gets an owner, people are trained in their new roles, and you receive the code and the documentation. ## Redrawn once. Kept current every month. Models get better, the business changes and new work turns up. After the ten weeks we can stay on monthly: we keep the agents tested and running, redraw the next part of the company and train people in roles that did not exist a year ago. - Next redesign: One more process or department redrawn and built each quarter. - Model updates: Agents retested and moved to a newer model when the tests show it does better. - Hours report: The hours that came back per department, measured against the baseline. - Role support: New people trained, and owners coached as their agents take on more. [Ask about the monthly plan](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Company%20Redesign%20care) ## What you receive. - The redrawn map: Every department’s steps before and after, with the hours that come back. - Working agents: Built on your stack, tested on your own records, handed over with the code. - A team that owns it: New roles written, people trained, an owner named for every agent. From $64,000 ten weeks or more, scoped to the size of the company An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For CEOs and leadership teams at 20–500 person companies. Useful when everyone has AI tools and the way the company runs has not changed. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Company%20Redesign) ## Where the redesign leads. It starts with a plan and builds from there. ### Floor Plan The first step: how the company works today, before anything is redrawn. **$6,400** ### Company Brain Usually the first build: the knowledge every later system reads. **$19,200** ### 100×worker Program Trains your people on the new way of working. **$8,000** ### Agent Ops We run the agents after handover, for as long as you want us to. **$3,200/mo** ## Before you ask. ### Do we need to have started with AI already? No. The redesign starts from how the work runs today, whatever tools are in it. Licenses and earlier pilots are useful input, and none are required. ### What happens to the people whose work changes? The map shows which roles change and how many hours come back before anything is built, so leadership decides with the numbers in front of it. We write the new roles and train people for them. Where employee representatives have to be consulted, we help prepare it early and in writing. ### Which models and tools do you build with? We are model-agnostic and usually start with Claude, built on the systems you already run. You own the context, the tests, the rules and the connections, so you can change models later. ### How is the price set? The price on this page is an indication from past projects. The number of departments and systems moves it, so we scope it on the map and agree it in writing before the build starts. A larger company can also split the work into phases. ### Can we start smaller? Yes. A Pilot Workshop designs one pilot in a day, and an AI Opportunity Map ranks where to start across the company. Both carry into a redesign later. --- # Eval Suite by VANDAM: Test every new model on your own work > A test suite built from your own work, a safe place to run it and one command that compares any model with the one in production. You see quality, cost and speed per case, and the switch becomes a decision with evidence behind it. URL: https://vandam.ai/eval-suite Contact: hendrik@vibelab.be # Know whether a new model helps before you switch. A test suite built from your own work, a safe place to run it and one command that compares any model with the one in production. You see quality, cost and speed per case, and the switch becomes a decision with evidence behind it. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Eval%20Suite) **$12,800** three weeks ## Most teams can watch their agents. Far fewer test them. - **32%** of 1,340 practitioners named quality as the top barrier to putting agents into production. [LangChain](https://www.langchain.com/state-of-agent-engineering), State of Agent Engineering. - **52%** run offline evaluations on test sets, while 89% have some form of observability for their agents. [LangChain](https://www.langchain.com/state-of-agent-engineering), same survey. - **62%** have tracing detailed enough to inspect individual agent steps and tool calls. [LangChain](https://www.langchain.com/state-of-agent-engineering), same survey. > At dinner, someone asked the table, “If a new model drops next month, does your company get better or worse?” HALF the table went quiet! > > Greg Isenberg, CEO, Late Checkout, October 2026, [Source](https://x.com/gregisenberg/status/2110414696037716372) So that dinner question gets an answer the same afternoon. ## Few companies can say how accurate their AI is. A new model every month makes that harder. An eval is a test for an agent: a real task, the answer your people would accept and a check that says right or wrong. Run them together and you know how well the agent does the job today, and whether a change makes it better or worse. Without them, every model release becomes a debate settled by a few tries in a chat window. [Alex Lieberman](https://x.com/businessbarista/status/2097478065723162959) summed up the recipe after a session on evals with LangChain: tasks, a verifier and an environment that looks real to the agent but cannot overwrite production. Each time a new model comes out, you run the suite and weigh cost, speed and intelligence against each other. In his list of thirty features of an AI-native company, [evals are core infrastructure](https://x.com/businessbarista/status/2094213970215231831). We build that suite for one to three of your agents. The tasks come from your own records. The rubrics are written with the people who do the work, because they decide what counts as right. Tracing is on, so a failed case shows the step where it went wrong. The suite never runs against live systems, and it stays with you: cases, rubrics and the runner sit in your repository, with the rest of [what stays yours](https://vandam.ai/models). The expert decisions in a [Judgment Vault](https://vandam.ai/judgment-vault) become cases here, and the reasoning is in [own your intelligence](https://vandam.ai/notes/own-your-intelligence). > In the world of agents, performance management is evals. > > Garrett Lord, CEO, Handshake, June 2026, [Source](https://x.com/GarrettLord/status/2068754262440767500) ## How the suite gets built. 1. Take the cases from the work: We pull real tasks from your records for each agent: requests, files, tickets and the answers your people sent. Edge cases go in on purpose, such as the order with no quantity or the discount just over the limit. 2. Write down what right means: Rubrics and verifiers are written with the people who do the work. Some checks are code, such as the price on the contract list. Others are a checklist that a second model or a person applies. Each check is one plain line. 3. Run it the same way every time: The suite runs in a test environment with copies of your systems and tracing switched on. One command runs any model against it and writes a report with quality, cost per task and speed side by side, case by case. ## Built once. Run on every model that matters. Models ship every few weeks and your work changes with them. After the handover we can run the suite on each release worth testing, turn production failures into new cases and tell you when a switch pays off. - New model runs: Each relevant release run against your suite within a week, with the report. - New cases: Failures from production turned into cases with the expected answer attached. - Cost and quality: Cost per task and pass rate per agent, tracked month by month. - Review: A monthly session with the rubric owners on what changed and what to switch. [Ask about model watch](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Eval%20Suite%20care) ## What you receive. - Task set: Real cases with expected answers, for one to three agents. - Rubrics and verifiers: The checks that decide right or wrong, one plain line each. - Model comparison: Quality, cost and speed per model, case by case. $12,800 three weeks An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For operations and technology leads with one to three agents, live or planned. Useful when a new model ships and nobody can say whether to switch. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Eval%20Suite) ## Where the suite leads. Each next step has its own service. ### Agent Ops We run the suite every month on the live agents, and before every model switch. **$3,200/mo** ### Token Ledger The cost side of every verdict: what the switch saves per task. **$9,600 + $1,600/mo** ### Private AI Tests open models you could run yourself against the ones you pay for. **$16,000** ## Before you ask. ### How many cases does a suite need? Enough to cover the work and its edge cases. For a focused agent that is often 40 to 200. We start with the cases that matter most and grow the set from production. ### We have no agent live yet. Is this too early? A suite written before the build tells the builders what right means, and it lets you compare models before you choose one. ### Which models can you compare? Any model you can reach through an API or run yourself, including open-weight models on your own hardware. We usually start with Claude and test the alternatives against it. ### Does this touch our live systems? No. The suite runs in a test environment with copies or simulations of your systems. Agents optimize for the score they are given, so they never practice on production. ### Who owns the suite? You do. Cases, rubrics, verifiers and the runner live in your repository and run without us. --- # Floor Plan by VANDAM: How your company works today: people, roles and workflows, mapped before any AI > Two weeks of interviews and shadowing across your departments. We draw the org chart as it actually works, every main workflow step by step and where the hours go. No AI yet: an honest picture of the company that every later decision builds URL: https://vandam.ai/floor-plan Contact: hendrik@vibelab.be # Map how your company works before anything changes. Two weeks of interviews and shadowing across your departments. We draw the org chart as it actually works, every main workflow step by step and where the hours go. No AI yet: an honest picture of the company that every later decision builds on. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Floor%20Plan) **$6,400** two weeks · credited if you continue ## Most change starts from a picture nobody checked. This one starts from how the work really runs. The org chart on the intranet says who reports to whom. It does not say who actually approves a discount, which spreadsheet the month-end close depends on, or why a new customer waits three days for a credit check. Those answers live with the people doing the work, and most projects start without asking them. A Floor Plan writes them down. We interview the people who do and manage the work in each department, sit with them for part of their day and read the documents, tools and inboxes the work runs through. Then we draw it: the departments and roles, every main workflow step by step, who owns each step, how long it takes and how long it waits before the next one. Our note on [mapping the work first](https://vandam.ai/notes/map-the-work-first) explains why this comes before any automation. It is deliberately not an AI project. The map holds whatever you decide next: an [AI Opportunity Map](https://vandam.ai/ai-opportunity-map) ranks opportunities on top of it, a [Company Brain](https://vandam.ai/company-brain) uses it as its spine, and a [Company Redesign](https://vandam.ai/company-redesign) redraws it. That is why every VANDAM project starts here, and why its fee is credited when you continue. ## How the plan is drawn. 1. Agree the scope: We agree which departments and workflows are in, and who we talk to. Usually every department, and the five to ten workflows that carry most of the hours. 2. Interview and shadow: Structured interviews with the people who do and manage the work, then time beside them while they do it. We collect the forms, spreadsheets, tools and inboxes each step runs through. 3. Draw the plan: The org chart as it works, each workflow step by step with its owner, time, tools and handoffs, and the hours each one takes in a week. 4. Check it with your team: Every department reviews its part. Corrections go in before anything is final, so the plan is the one your people recognise. ## What you receive. - Org chart as it works: Departments, roles and people, with who decides what and who depends on whom. - Workflow maps: Your main workflows step by step, with the owner, the time, the tools and every handoff. - Where the hours go: Hours per workflow, and the waiting, rework and bottlenecks between the steps. $6,400 two weeks · credited if you continue An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For leadership teams at 20–500 person companies. The first step of every VANDAM project. If you continue with another service, the fee is credited against it. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Floor%20Plan) ## Where the Floor Plan leads. Each next step builds on it. ### AI Opportunity Map Ranks the AI opportunities on top of the plan: which workflows to change first, and how. **$9,600** ### Agent-Ready Audit Checks whether the data and systems in the plan can carry agents, and what to fix first. ### Company Brain Uses the plan as its spine: departments, roles and people first, every other page hangs off them. **$19,200** ### Pilot Workshop Picks one workflow from the plan and designs it as a pilot with your leadership team in a day. **$4,800** ### Company Redesign Redraws the whole plan around agents and people, and builds the services in order. **From $64,000** ## Before you ask. ### Is there any AI in it? No. The Floor Plan describes how the company works today. Where AI could help comes after, in an AI Opportunity Map or a Company Redesign, and both build on this plan. ### How is the fee credited? If you start another VANDAM service within three months, the Floor Plan fee is deducted from its price. ### How much of our team’s time does it take? About an hour of interview per person we speak to, some time beside a few people while they work, and a review of their part of the plan at the end. We agree the list with you before we start. ### Do we need VANDAM after it? No. The plan is yours and stands on its own. Many teams use it to scope work with their own people or other partners. ### What if our company is small? Then the plan is smaller and so is the price. Below about 30 people, two departments and their workflows can often be drawn in one week. --- # Invoice to ERP by VANDAM: Supplier invoices matched and posted to your ERP > A system that reads supplier invoices, matches their lines to your articles and prices, and posts under your approval rules. Your finance team receives a review queue for the items that need a decision. URL: https://vandam.ai/invoice-to-erp Contact: hendrik@vibelab.be # Supplier invoices checked. Exceptions ready for review. A system that reads supplier invoices, matches their lines to your articles and prices, and posts under your approval rules. Your finance team receives a review queue for the items that need a decision. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Invoice%20to%20ERP) **$24,000** three weeks ## Supplier payments are where the fraud goes. - **$3.05B** lost to business email compromise in 2025, a scam the FBI describes as targeting businesses that work with suppliers and pay by wire. [FBI IC3](https://www.ic3.gov/AnnualReport/Reports/2025_IC3Report.pdf), 24,768 complaints. - **74%** of US organizations were hit by business email compromise in 2025. Only 17% use AI against payments fraud. [AFP](https://www.financialprofessionals.org/about/learn-more/press-releases/Details/over-75-percent-of-us-firms-experienced-payments-fraud-in-2025-while-ai-adoption-for-fraud-mitigation-lags) survey of 465 treasury professionals, 2026. - **$9.90** average cost to process one invoice, against $2.67 for the best teams, with one in five invoices raising an exception. [Ardent Partners](https://www.medius.com/resources/guides-reports/ardent-partners-accounts-payable-metrics-that-matter/), 2026, published with an AP software vendor. > Invoices are in Quickbooks, emails are in Outlook, proposals live in Sharepoint, contracts live in Netsuite, and so on. There is no shared namespace to access all this data across the business. > > Eli Mernit, Founder, Beam, February 2026, [Source](https://x.com/mernit/status/2021324284875153544) So no, a smarter OCR alone won’t fix this. The matching rules do. ## Give Finance the differences that matter. An invoice arrives as a PDF or a scan. The article list and the agreed prices live in the ERP. Someone sits between the two, retyping lines and chasing anything unfamiliar, and the routine lines arrive in the same pile as the difficult ones. Most of that pile needs rules, and a small part needs a person. [Annelies Gamble](https://x.com/AnneliesGamble/status/2069814714214023186), a partner at Zetta Venture Partners, draws the line in her June 2026 essay on agent deployments. “Deterministic automation is for rules,” she writes, with routing an invoice as her example. “Humans are for accountability, ambiguity, and trust.” We build to that split. Matching lines post under the rules you approve. A price difference or an unknown item goes to a review queue with both records side by side and the reason it was flagged, so Finance spends its time on decisions instead of retyping. Invoice to ERP builds on your [company brain](https://vandam.ai/company-brain): each invoice is checked against the agreements and price lists kept in it. ## From supplier document to accounting record. 1. Read your actual invoices: We configure the document reading around PDFs and scans from your suppliers. Article references, quantities and prices keep their source location so the team can check an extracted field. INV-20931.pdf · fields read - **Supplier** Brightline Packaging - **Article** BX-440 carton, 40×30 - **Quantity** 1,200 · page 1, line 3 - **Unit price** $0.84 · page 1, line 3 - **Total** $1,008.00 2. Match and show exceptions: We compare the lines with your supplier records and price agreements. Unknown items and differences enter a review queue with both records visible and a clear reason for the check. Review queue · 1 of 4 Invoice Price agreement - BX-440 carton BX-440 carton - 1,200 units 1,200 units - $0.84 each $0.79 each Price differs by 6%. Agreed tolerance is 2%. 3. Connect the approved posting: We connect the system to your ERP under the agreed approval rules. Your team receives editable matching rules and a way to monitor changes, quantity differences and duplicate documents. Posted to ERP today - 86 invoices matched and postedauto - 4 waiting in reviewFinance - 1 duplicate blockedINV-20877 - Matching rules changed0 ## Live in three weeks. Looked after every month. You own the rules, the code and the connection. New suppliers, new document layouts and price changes keep arriving, so after go-live we can stay on for a monthly fee. - Hosting: We run the document reading, the matching and the ERP connection, on your infrastructure or ours. - New suppliers: New suppliers and document layouts set up as they arrive, so they match from their first invoice. - Rule review: A monthly look at what reached the queue, and which corrections should become rules. - Check-ins: A monthly session with Finance on match rates, open exceptions and what to automate next. [Ask about monthly care](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Invoice%20to%20ERP%20care) ## What you receive. - Document extraction: Configured for your supplier documents. - Matching rules: Your suppliers, article codes and prices, editable by your team. - Review queue: Exceptions with the source and reason shown together. - ERP connection: Posting into the system you already use. supplier lines matched to articles for a nine-site retail group, against their live ERP. $24,000 three weeks An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For finance and operations teams at multi-site businesses. Useful when invoices are retyped or checked against separate price sheets. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Invoice%20to%20ERP) ## Where Invoice to ERP leads. Each next step has its own service. ### Company Brain Comes before: the agreements and price lists every invoice is checked against. **$19,200** ### Agent Ops We run it after go-live: every run logged, every failure fixed and explained. **$3,200/mo** ### Token Ledger What the agent costs per invoice, next to the hours it saves. **$9,600 + $1,600/mo** ### Eval Suite Tests every new model on your own invoices before it touches the ledger. **$12,800** ## Before you ask. ### Does it post anything without approval? Only lines that match your records within the tolerances you set. Everything else waits in the review queue for a person. ### Which ERP systems does it work with? Systems with an API or an import format. We confirm the connection for your ERP before the project starts. ### What happens with scanned or handwritten invoices? Scans are read like PDFs. Fields the system cannot read with confidence go to the queue with the original image next to them. ### Can Finance change the rules themselves? Yes. Rules are written in plain words and every change is logged with who made it. --- # Judgment Vault by VANDAM: Expert decisions, documented and tested > A documented decision guide built from one person’s experience. Your team receives the rules, reference cases and explanations behind recurring decisions, with clear boundaries for when to act and when to ask. URL: https://vandam.ai/judgment-vault Contact: hendrik@vibelab.be # Your expert’s decisions. A guide others can use. A documented decision guide built from one person’s experience. Your team receives the rules, reference cases and explanations behind recurring decisions, with clear boundaries for when to act and when to ask. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Judgment%20Vault) **$16,000** four weeks · per decision maker ## The best judgment in the company sits with a few people. - **15%** more issues resolved per hour when 5,172 support agents got an AI assistant trained on their top performers. The least experienced gained most. [NBER](https://www.nber.org/papers/w31161), Brynjolfsson, Li and Raymond. - **11,200** Americans turn 65 every day from 2024 through 2027, more than 4.1 million a year. [Peak 65 research](https://www.planadviser.com/peak-65-finds-millions-americans-unprepared-economist-warns/), Alliance for Lifetime Income. > This is the actual bottleneck. The models are smart enough already. What is missing is the company-specific context locked in senior people heads. > > Garry Tan, President and CEO, Y Combinator, May 2026, [Source](https://x.com/garrytan/status/2061076595474121073) So no, waiting for the next model won’t help. Your expert’s past cases will. ## Keep the reasoning in the business. Some decisions keep landing on one desk: a price exception, a credit request, a supplier nobody has used before. The person at that desk knows which two details matter. Everyone else only sees the answer, so the next similar request starts the same conversation again, and the work waits whenever that person is away. We sit with that person and go through real past decisions until the reasoning is on paper: what they checked, what made the exception different, where they would refuse. Part of the job is cutting. Some habits are real judgment and some are workarounds for a system that broke years ago, and only the first kind belongs in the guide. What remains becomes a guide colleagues can read and instructions an AI system can follow within agreed limits. The unusual cases still go to the expert. If you just capture it and encode it into the agent, you’ve automated the dysfunction. (Ahmad Kakar, CEO of Euclid, quoted by [Annelies Gamble](https://x.com/AnneliesGamble/status/2069814714214023186) of Zetta Venture Partners, June 2026) Judgment Vault builds on your [company brain](https://vandam.ai/company-brain): the expert’s decisions become rulings in it, next to the policies they apply. ## From experience to a working guide. 1. Review real decisions: We interview the decision maker using previous cases. We record the information considered, the reason for the answer, and what made an exception different from an ordinary request. Case 041 · credit request - **Customer** Harbor Foods, 6 years - **Asked** $120k limit, up from $80k - **Looked at** Payment days, order trend, owner change - **Decision** Approved at $100k - **Why the exception** Pays in 21 days; the new owner is the old CFO 2. Write and test the guide: We turn that material into rules, examples and escalation points. Then we compare the guide against known decisions and review disagreements with the expert before it is used. Guide vs. known decisions Guide says Expert decided - Approve $100k Approved $100k - Decline Declined - Escalate Approved $60k 38 of 40 match. 2 under review. 3. Introduce it with supervision: Your team starts with the guide under agreed supervision. You receive the decision library, test cases and ownership record, including the circumstances in which someone must ask for judgment. Supervised use, week 2 - Requests under $50kteam decides - $50k to $150kguide + check - Above $150kask the expert - New customer, no historyask the expert ## Written in four weeks. Kept current after that. You own the guide, the cases and the skills. Customers, prices and the expert’s own thinking keep changing, so after handover we can stay on for a monthly fee to keep the guide and its tests in step. - New cases: Requests that went to the expert are added as reference cases, so the guide learns from them. - Retesting: Every rule change is run against all reference cases before your team uses it. - Boundary review: The escalation points are reviewed against the record, and they move only when the cases support it. - Expert sessions: A monthly session with the expert on disagreements, new exceptions and decisions to add next. [Ask about monthly care](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Judgment%20Vault%20care) ## What you receive. - Decision library: Readable guidance and tested agent skills. - Reference cases: Past decisions retained as repeatable tests. - Decision record: The interviews, reasoning and ownership of each rule. $16,000 four weeks · per decision maker An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For owners and department heads whose decisions depend on one person. Useful before a handover, or when routine decisions keep returning to the same desk. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Judgment%20Vault) ## Where the vault leads. Each next step has its own service. ### Company Brain Comes before: the policies and history the rulings sit next to. **$19,200** ### Agent Onboarding Agents load the rulings for the decisions they help with. **$9,600 + $1,600/mo** ### Eval Suite Tests agents against your expert’s reference cases on every new model. **$12,800** ### 100×worker Program Trains the next people on the decisions the vault holds. **$8,000** We can know more than we can tell. (Michael Polanyi, [The Tacit Dimension](https://www.taylorfrancis.com/books/9780203559055/chapters/10.4324/9780203559055-8), 1966) ## Before you ask. ### How much of the expert’s time does this take? The interviews use real past cases, and the expert reviews every disagreement from the test run. We agree the sessions at the start so they fit around the expert’s normal work. ### Will an AI make these decisions on its own? Only within the boundaries agreed during the supervised start. Unusual cases and anything past an escalation point still go to the expert. ### What happens when the guide and the expert disagree? Every disagreement is reviewed with the expert before the guide is used. It ends as a new rule, a new exception or a corrected case, and the decision record notes which. ### Can we cover more than one person? Yes. Each decision maker is a separate project with their own interviews, cases and tests. --- # Local Law 144 Audit by VANDAM: Bias audits for hiring tools covered by NYC Local Law 144 > An independent bias audit, a public results summary and candidate-notice drafts for hiring tools covered by NYC Local Law 144. The engagement starts by establishing which tools and uses are within scope. URL: https://vandam.ai/local-law-144-audit Contact: hendrik@vibelab.be # A bias audit. The publication pack to go with it. An independent bias audit, a public results summary and candidate-notice drafts for hiring tools covered by NYC Local Law 144. The engagement starts by establishing which tools and uses are within scope. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Local%20Law%20144%20Audit) **$8,000** annual · repeats every year ## Few employers publish the audit, and few get checked. - **18/391** New York City employers studied had posted a bias audit. Thirteen posted the required notice. [Cornell](https://www.arxiv.org/abs/2406.01399) and Data & Society, 2024. - **1:17** The city reviewed 32 companies and found one issue. State auditors reviewing the same companies found at least 17. [NY State Comptroller](https://www.osc.ny.gov/state-agencies/audits/2025/12/02/enforcement-local-law-144-automated-employment-decision-tools), December 2025. - **$1,500** maximum penalty per repeat violation, and every day a tool runs out of compliance counts as a separate one. [Local Law 144](https://intro.nyc/local-laws/2021-144). ## Bring the tool, the audit and the notices together. Hiring software can play different roles in a decision, so the first question is how your organization actually uses it. For covered automated employment decision tools, [New York City](https://www.nyc.gov/site/dca/about/automated-employment-decision-tools.page) has enforced Local Law 144 since July 5, 2023. The tool needs a bias audit dated within **one year** of its use, a published summary of the results, and a notice to candidates and employees **10 business days** before use. We assemble that work into one traceable file for your HR and legal teams. The audit ties its results to the data and method behind them. The publication pack holds the summary and notice drafts for your review, and records where and when each one has to appear. ## How the audit file is prepared. 1. Confirm the tool and use: We inventory the relevant tools and document their role in hiring or promotion. The scope assessment determines which uses are covered and defines the work needed for the engagement. Tools in hiring - Resume screener · ATScovered - Interview schedulernot covered - Assessment test scoringcovered - Promotion reviewnot used 2. Perform the independent audit: The auditor documents the data, method and required selection or scoring rates and impact ratios. The results include the applicable category and intersectional breakdowns, with the supporting calculations retained. Impact ratios · example - Category A1.00 - Category B0.91 - Category C0.84 - Category D0.78 3. Prepare the publication materials: We prepare a dated public summary and the applicable candidate-notice drafts. Your team receives the results, supporting references and a record of required timing and placement for review and publication. Publication file - Public summary, datedready for review - Candidate notice, 10 business daysdraft - Careers page placementagreed - Next audit duewithin a year ## Dated once. Due again within the year. The city requires a bias audit within one year of the tool’s use, so a covered tool that stays in use needs a newer audit every year. The engagement repeats annually, and the file, method and data requests carry over. - Yearly audit: New data, the same method, and a new dated summary and notice drafts. - Scope review: When a tool, its version or its use changes, we check whether the scope assessment still holds. - Audit calendar: The audit date, posting dates, first use and one-year mark tracked for each tool. - HR and legal check-in: A session before each audit on tool changes, data and placement. [Ask about the yearly audit](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Local%20Law%20144%20Audit%20care) ## What you receive. - Tool inventory: The tools, their uses and the scope assessment. - Bias audit: Category and intersectional results, with the required impact ratios. - Public summary: A results summary prepared for publication. - Candidate notices: Drafts and a record of where they need to appear. $8,000 annual · repeats every year An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For HR, people and legal teams using automated hiring tools. Start by confirming which tools and uses fall within the law’s scope. [NYC requirements](https://www.nyc.gov/site/dca/about/automated-employment-decision-tools.page) [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Local%20Law%20144%20Audit) ## Where the audit leads. Each next step has its own service. ### Eval Suite Tests the hiring tool again whenever its model changes, between yearly audits. **$12,800** ### Source Cleanup Cleans the candidate data the tool reads, where the audit traced bias to it. ### Agent Attack Audit Red-teams the screening agents and what they can reach. **$6,400** DCWP’s AEDT complaint process is ineffective in ensuring that all complaints related to non-compliance with LL144 are routed to DCWP. ([New York State Comptroller](https://www.osc.ny.gov/state-agencies/audits/2025/12/02/enforcement-local-law-144-automated-employment-decision-tools), audit of city enforcement, December 2025) ## Before you ask. ### Does Local Law 144 apply to our tools? That depends on how each tool is used in hiring or promotion decisions, which is why every engagement starts with a scope assessment. We document the use and the reasoning, and your legal team makes the final call. ### What data do you need from us? Records of the tool’s use, with the outcomes and the sex and race or ethnicity categories applicants reported. Applicants with an unknown category are counted and shown separately. ### What if an impact ratio is low? Local Law 144 sets no pass mark. The audit reports the results, and what they mean for your hiring is a question for your legal team, since other anti-discrimination laws still apply. ### Is this legal advice? No. We deliver the audit, the summary and the notice drafts. Your HR and legal teams review them and decide what is published, where and when. --- # MCP Retrofit by VANDAM: Connect your product to Claude, ChatGPT and other MCP clients > A working connection between your software and customers’ AI assistants. They can ask for agreed actions, such as finding an order, while your product continues to control which accounts, records and changes they may access. URL: https://vandam.ai/mcp-retrofit Contact: hendrik@vibelab.be # Let customers use your product from their AI assistant. A working connection between your software and customers’ AI assistants. They can ask for agreed actions, such as finding an order, while your product continues to control which accounts, records and changes they may access. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20MCP%20Retrofit) **$12,800** two weeks · plus hosting ## Within a year, every major assistant spoke MCP. - **10,000+** active public MCP servers by December 2025, when [Anthropic](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation) handed the protocol to a Linux Foundation fund backed by OpenAI, Google, Microsoft and AWS. - **97M** monthly downloads of the MCP software kits at that point, used inside ChatGPT, Gemini, Copilot, Cursor and VS Code. [Anthropic](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation). - **4.5** months from Anthropic’s launch in November 2024 to support from [OpenAI](https://x.com/sama/status/1904957253456941061) and Google DeepMind. > MCP is emerging as a new AI interface. In the near-future, MCP may become the default way, or in some cases the only way, people, businesses, and code discover and interact with services. > > Jeff Weinstein, Product lead, Stripe, May 2025, [Source](https://blog.cloudflare.com/mcp-demo-day/) So no, your login page won’t stay the only way into your product. ## Make your product available where work begins. A customer who already works in an AI assistant wants to find an order or check stock by asking there, without opening your product in another tab. If your software cannot answer, the assistant guesses or sends them elsewhere. [Anthropic](https://www.anthropic.com/news/model-context-protocol) released MCP, the Model Context Protocol, as an open standard in November 2024, and [Block](https://www.anthropic.com/news/model-context-protocol) was among the first companies to build on it. It lets an assistant call defined actions in your software. We build that connection around a small, agreed set of actions, and your product keeps its own permissions: a customer who can read orders still cannot change one because the request came through an assistant. ## How the connection is delivered. 1. Choose the useful actions: We agree which tasks customers should be able to complete from their assistant. Each action gets a clear description, required information and an explicit access boundary tied to your product. Actions for assistants - Check order statusany signed-in customer - Book a delivery slotaccount owner - Download an invoicebilling role - Change pricesnever exposed 2. Build and test the connection: We build the MCP service, authenticate customers and apply account permissions. Tests cover complete requests, disallowed actions and failure cases, including how an assistant should explain an unsuccessful request. Test run · 48 cases - Complete requests22 passed - Disallowed actions refused14 passed - Failure explained to the user11 passed - Expired token1 fixed, retest 3. Launch with operating instructions: We deploy the service on your infrastructure or ours. You receive the running connection, action documentation, repeatable tests and the instructions your technical team needs to operate and maintain it. atlas-mcp · deployed - tools: 3 auth: OAuth 2.1 - clients: Claude, ChatGPT, Copilot - GET /orders/{id}/status - POST /deliveries/slots - GET /invoices/{id}.pdf ## Live in two weeks. Looked after when you want. You own the service, the tests and the documentation. Customers will ask for new actions and your product will change, so after launch we can stay on for a monthly fee. - Hosting: We run the MCP service on your infrastructure or ours. - Test runs: The conversation tests run again after each change to your product or to the actions. - New actions: New actions scoped, built and tested with your product team, a few at a time. - Check-ins: A monthly session with your product team on open requests and which action to add next. [Ask about monthly care](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20MCP%20Retrofit%20care) ## What you receive. - MCP server: Customer-specific keys, authentication and rate limits. - Conversation tests: Repeatable checks for the actions your users take. - Tool documentation: Action descriptions, parameters and access requirements. - Deployment: A running service and its operating guide. MCP servers in production: Douano, Marketing, Kompas and NextCV. Douano runs orders and stock on a live sales stack. $12,800 two weeks · plus hosting An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For product and engineering teams at B2B software companies. Useful when customers want to use your product from their AI assistant. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20MCP%20Retrofit) ## Where the retrofit leads. Each next step has its own service. ### Agent Attack Audit Red-teams the new interface before customers’ agents use it. **$6,400** ### Agent Relationship Audit How assistants find, compare and recommend you once they can act on your product. **$11,200** ### Agent Ops We keep the interface running and answer when a client changes how it calls you. **$3,200/mo** > Every B2B software company is (or should be) building a "headless" version of their product. One that can be used by agents. > > Dharmesh Shah, Co-founder and CTO, HubSpot, April 2026, [Source](https://x.com/dharmesh/status/2043046723950436368) ## Before you ask. ### Can an assistant do more than the customer can in our product? No. Every call runs under the customer’s own account and permissions. A customer who can read orders still cannot change one through an assistant. ### Do we have to expose our whole API? No. We start from a small, agreed set of actions. Anything outside that list, such as changing prices, is never exposed. ### Which assistants can connect? Claude, ChatGPT and other clients that support MCP. Customers connect with their own account in your product. ### Where does the service run, and who maintains it? On your infrastructure or ours. You receive the operating guide and the tests, so your technical team can run it, or we can look after it monthly. --- # Models · VANDAM > The AI models VANDAM builds with: Claude, ChatGPT, Gemini, Copilot, Grok and open-source models, chosen per task and around the tools your team already uses. URL: https://vandam.ai/models Contact: hendrik@vibelab.be # The models we work with. We build with the model that fits the task, your rules for data and the tools your team already pays for. - Claude: Anthropic Long documents, careful writing and agents that work with your tools. - ChatGPT: OpenAI The assistant many teams already use, with custom GPTs and a broad API. - Gemini: Google Inside Google Workspace: Gmail, Docs, Drive and Meet. - Copilot: Microsoft Inside Microsoft 365: Outlook, Word, Excel and Teams. - Grok: xAI Answers that draw on live posts from X. - Open source: Llama, Mistral, DeepSeek, Qwen Runs on your own servers when information cannot leave the company. ## What stays yours Most companies rent their AI: they send their work to a lab and get answers back. We build so that everything around the model belongs to you and lives in your systems. - **Your context** The documents, decisions and exceptions the agents read, kept as files in your own storage. - **Your evals** Tests written with your own experts, so every model has to prove itself on your work. - **Your rules** Instructions, permissions and review steps, written down and owned by a named person. - **Your connections** Access to your systems through MCP, which works with any model that supports it. With those four in place, moving to a newer model, or to an open-weight model on your own servers, means running your evals again. [Why this matters now](https://vandam.ai/notes/own-your-intelligence) ## What we suggest first VibeLab, the team behind VANDAM, is a Claude implementation partner, and Hendrik is a Claude Ambassador. Claude is the stack we know best, so for a new build it is usually what we suggest first. If your company runs on Microsoft 365, Google Workspace or ChatGPT, or has a stack it prefers, we build on that. Agents reach your systems through MCP, an open standard, so a later change of model keeps the connections in place. ## How we choose 1. Your task: We try representative questions, documents and actions. Your team reviews the result against what a useful answer should contain. 2. Your environment: We consider the systems you use, access requirements and where your information can be processed. 3. Everyday operation: We compare response quality, speed and cost for the agreed workload, and document what should be reviewed as the work changes. **Not sure which fits?** The AI Opportunity Map compares them on your own work. [AI Opportunity Map](https://vandam.ai/ai-opportunity-map) --- # Notes · VANDAM > Notes on AI diffusion, org design and how companies change, with every number linked to its source. URL: https://vandam.ai/notes Contact: hendrik@vibelab.be # Notes on how companies change. What we read, measure and learn while redesigning companies around AI. Every number links to its source. 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. ## Fix a mistake once by turning every correction into a rule the agents read. Anthropic's data team took Claude from 21% to above 95% accuracy with written skills, and now turns every correction a colleague makes into a rule and a test for the person who owns that domain. I set out that loop for a mid-sized company, with where the rules live, who owns them, and the one number a COO should track: how often the same mistake comes back. ## Plan for the week your AI vendor goes away, before someone else picks the date. In February 2026 defense contractors started moving off Claude within days, for reasons unrelated to the model, and every provider now publishes retirement dates for the models you depend on. Why the switching cost sits in prompts, evals, traces and data access, and the supplier-risk checklist a CFO and COO can run before the next renewal. ## Most of your AI list needs software and only a few items need an agent. Agents still finish a minority of office tasks and give different answers on repeated runs, while a written rule does the same thing every time for almost nothing. I set out four questions that sort every request into software, agent or person, and apply them to ten typical ones. ## Your outside firms got faster with AI and you are still paying for the old hours. Law firms, auditors, agencies and outsourcers are putting AI into the work they bill you for, and most contracts still price the old hours. KPMG showed how to ask for the saving back, and I set out the questions a CFO should send each supplier before the next renewal. ## Check every invoice, not one in twenty, now that a judgment costs a tenth of a cent. Sampling exists because reading every document used to cost a person’s time, and a small model now reads an invoice with its PO and receipt for less than a tenth of a cent. I work through the arithmetic, what the PCAOB and the ACFE say about whole populations, and a monthly plan and budget for a 500-person company. 26 October 2026 8 min read ## The work you never did because it never paid is where the largest AI return sits. Most AI business cases count the hours of work a company already does, which caps the return at today's labor cost. The evidence from Jevons to bank tellers says cheaper work expands where demand was rationed, and I set out how a leadership team lists, prices and picks the expert work it skips. ## Ask what your agents can reach because any text they read can give them orders. Prompt injection tops OWASP’s list of LLM risks, and every frontier model in a NIST-run red-teaming contest was hijacked at least once, so the limit has to sit in what an agent can touch. I lay out the one-page reach review we use and how to give employees who bring their own agents a sandbox instead of a ban. 25 October 2026 8 min read ## Why most AI pilots stall and how to choose one that scales. MIT's famous 95% figure measures less than the headline says, but larger surveys from S&P Global, Gartner and McKinsey point the same way. The pilots that reach production share five conditions, and a CEO can check for them before the first demo. ## Measure the work before you let AI change it. Eight in ten survey respondents say AI makes them more productive, yet only about a third of companies can find it in EBIT, and few measured the work before it changed. Why the companies with credible AI numbers all name a starting point, and a three-week baseline method a CFO can run before the next AI purchase. ## Before you automate anything, find out where the hours actually go. Most AI shortlists are built from the visible work, while the hours sit in coordination, waiting and exceptions that no process chart shows. The evidence from 1990 to 2026, and the two-week work map we use to rank tasks by hours times feasibility. 25 October 2026 9 min read ## What managers do now that part of every team is agents. Amazon, Microsoft and Meta cut management layers while agents took on coordination work, and the managers who remain carry wider spans and a second team to supervise. What Gallup and Anthropic's data say about that job, and a manager role redesigned around it. 25 October 2026 7 min read ## Once agents do the middle of the work, your people should own both ends. Balaji Srinivasan’s line that AI works middle-to-middle is the most useful frame I know for redesigning a role. This note covers what sits at the start and the end of a task, what 1,500 workers told Stanford they want, and how I would rewrite one production manager’s morning. ## How to tell your people what AI will change about their jobs. Most employees can see AI arriving, and only a quarter say their employer has explained a plan. What silence and shifting explanations cost, a communication sequence for the CEO and CHRO, and what never to promise. ## The AI line in your budget should be priced per finished task. AI spend is moving from seats and headcount to tokens and outcomes, while most finance teams still count licenses. Intercom, Box, Podium and Varick show what the cost side of AI looks like, and I name the three numbers I would ask a CFO to own. ## AI spreads through a few people who build for their own work. Shopify and Duolingo both told every employee to use AI, and Duolingo dropped its usage metric a year later. The people who actually spread AI build tools for their own work, and a company can give them time, status, budget and a way to share what works. ## What changes in your controls when an agent can post to the ledger. Once an AI agent posts entries, approves invoices or touches vendor records, it needs a place in the access matrix, a change process and a human who signs for it. What COSO, the PCAOB and the audit firms actually say, and the checklist we use for an agent in accounts payable. ## What a firm should keep inside once agents make coordination cheap. In 1937 Ronald Coase explained that firms exist because using the market costs money. Agents now make buying, building and internal coordination cheaper all at once, and I list the five comparisons I would rerun at a 300-person company. ## AI transformation has ten steps and only one of them is AI. Alex Lieberman of Tenex listed the ten steps it takes to make a process AI-native. I walked them through one quote process at a parts distributor to show where the weeks go, and why the map pays off before any model does. 24 October 2026 7 min read ## Europe can lose the model race and still win on AI. At a keynote this week I heard the case that Europe should stop chasing frontier models and put everything into getting AI used in every company. I checked the evidence behind it, read the strongest objection, and worked out what it means for a single firm. ## Who should do your AI work, now that the labs sell it too. In 2026 Anthropic and OpenAI set up services companies, Sequoia and Y Combinator told founders to sell the work, and Accenture bought an AI firm. A mid-sized company now chooses between a product, a big consultancy, a lab-backed team and a small senior team. 24 October 2026 6 min read ## Who trains your next senior once agents do the junior work. Four studies from the past year point the same way: AI is thinning the entry-level rung, mostly by slowing hiring. The judgment employers now look for was learned on that rung, so I set out what I would change as an employer. ## What your company knows becomes the asset once agents do the work. Agents only know what someone wrote down, and most companies record the outcome of a decision while losing the reason, the exception and the person who said yes. This note explains why that record compounds and lists the four things I would start capturing this month. ## Why AI licenses rarely make a company faster. Most companies now count AI logins. The few that get value from it changed the work itself first, the way factories once had to rebuild around the electric motor. 24 October 2026 8 min read ## What Anthropic found when its agents outworked its people. Anthropic measured its own company after agent hours passed human hours six to one, and the findings are about attention, management, hiring and people moving. I read the report as a story and draw out what it means for a company of 300. ## The new org chart has written rules. Block, Coinbase and Fiverr published how they are reorganizing around AI. The rules are concrete, most of them came with layoffs, and the people side decides whether they work. ## Why our AI team hires from HR and finance. Most AI projects speed up the process a company already has, and redesigning it takes people who know where the money and the jobs sit. One month-end close, walked step by step, shows why they belong next to the engineers and why VANDAM is hiring them. --- # Why AI licenses rarely make a company faster. · VANDAM > Adoption counts logins, diffusion changes the work. Why 88% of firms use AI, 5% get value, and what redesigning the work looks like inside one company. URL: https://vandam.ai/notes/adoption-vs-diffusion Contact: hendrik@vibelab.be # Why AI licenses rarely make a company faster. Most companies now count AI logins. The few that get value from it changed the work itself first, the way factories once had to rebuild around the electric motor. Hendrik De Winne 24 October 2026 6 min read A leader of an operations organization with thousands of people told [Vas (@vasuman)](https://x.com/vasuman/status/2085806422072418632), the CEO of [Varick Agents](https://varickagents.com), that his team had been moving at a decent rate before AI. Then they rolled out Claude Cowork. In Vas’s words, “the team kept moving at the same rate.” He describes the usage behind that flat line as a barbell. **5‑10%** of the people became power users, with skill files and connectors to Outlook. Of the rest, **20%** used it a couple of times a day, “pretty poorly.” The remaining **70%** did not use it at all. It is one company, told by someone who sells agents for a living, so I read it as an observation and not as a study. I still recognize the shape of it from most companies I walk into. > The dashboards in his org would indicate that the rollout counts as adoption. But to him, nothing got faster. Both of these were true simultaneously. > > Vas, CEO, Varick Agents, August 2026, [Source](https://x.com/vasuman/status/2085806422072418632) Both of those were true at the same time because they measure different things. A dashboard counts who logged in. The leader was asking whether the work moved faster. Most companies I meet have bought the first and are waiting for the second to follow by itself. ## Every measure shows the same gap The pattern holds at much larger scale. In March, [Anthropic](https://www.anthropic.com/research/labor-market-impacts) compared what language models could do in theory with what people actually use Claude for. In computer and math jobs, the occupations closest to the technology, models could take on **94%** of the tasks. Claude covered **33%**. Could do 94% Actually used 33% Tasks in computer and math jobs. Source: [Anthropic](https://www.anthropic.com/research/labor-market-impacts) , March 2026. That is the group with the most technical people, the most curiosity and the least fear of the tools. The capability is already there. The space between the two bars is made of habits, handoffs, approval steps and job descriptions written before any of this existed. The surveys of companies say the same thing from the other end. [McKinsey](https://www.mckinsey.com/capabilities/operations/our-insights/the-state-of-ai) found that **88%** of organizations use AI regularly in at least one business function. [BCG](https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap) puts the share of firms getting value from it at scale at **5%**, and says 60% are reaping hardly any material value at all. - **88%** Organizations that use AI, according to [McKinsey](https://www.mckinsey.com/capabilities/operations/our-insights/the-state-of-ai), 2025. - **5%** Firms getting value from AI at scale, according to [BCG](https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap), 2025. - **3×** How much more likely high performers are to have fundamentally redesigned their workflows, per [McKinsey](https://www.mckinsey.com/capabilities/operations/our-insights/the-state-of-ai). The third number is the one I care about. When McKinsey looked at what separates the small group of high performers from everyone else, one of the clearest differences was that they were nearly three times as likely to have fundamentally redesigned their workflows. They bought the same models as everyone else and kept going until the work itself looked different. These are self-reported surveys, and I hold them loosely. But the direction never changes, whichever firm runs the survey and however it defines value: wide use, thin results, and a small group that rebuilt how the work flows. ## Factories took decades to learn this once The economic historian Paul David described the same gap in the electrification of factories, in [The Dynamo and the Computer](https://gwern.net/doc/economics/automation/1989-david.pdf). When electric motors arrived, many factories took out the steam engine and put an electric one in its place. The line shafts and belts that ran the length of the building stayed where they were, and so did the layout of the floor, the order of the work and the jobs of the people standing next to the machines. Electricity was in the building, and productivity barely moved. The gains came when factories were redesigned around what electricity made possible: a small motor on each machine, which meant machines could be placed where the flow of work wanted them instead of where the shaft ran. That change was slow, because it meant new buildings, new layouts and new ways of organizing people. The technology had been ready for a long time before the companies were. A chatbot added to an unchanged workflow is the dynamo bolted to the old line shaft. Everyone has access, the invoice still passes through the same inboxes, and the dashboard reports adoption. > The best way to get real efficiency gains is you have to actually reengineer the process to take advantage of what agents are actually good at and can do differently and faster than people. > > Aaron Levie, CEO, Box, September 2026, [Source](https://x.com/levie/status/2096279756925026733) ## Historians and central bankers call it diffusion Jeffrey Ding, a political scientist at George Washington University, made the national version of this argument in [Technology and the Rise of Great Powers](https://press.princeton.edu/node/51524). He looks at Britain in the first industrial revolution, America and Germany overtaking Britain in the second, and Japan’s challenge to America in the third. The countries that pulled ahead were the ones better at adapting new technologies at scale, and his theory stresses “institutional adaptations oriented around diffusing technological advances throughout the entire economy.” Invention mattered less than spreading it. The European Central Bank now says something close to that about AI. In a speech in March, [Philip Lane](https://www.ecb.europa.eu/press/key/date/2026/html/ecb.sp260323_1~1e06784a89.en.html), a member of the ECB’s Executive Board, said the aggregate productivity impact of AI “may vary widely, primarily driven by the speed of diffusion,” and that “firm adoption and organisational restructuring are historically the main channels” through which technologies like this have paid off. When a central banker and a historian of great powers land on the same variable, I take it seriously. Both point at how fast ordinary companies change the way they work, whoever ends up with the best model. That is the distinction I use with every client. Adoption is people having access to a tool, and you can buy it in a quarter. Diffusion is the work itself changing shape around the tool: who does which step, what gets checked by whom, which approvals disappear, which roles grow. That part only the company itself can do, one workflow at a time. ## Inside one company: map, redesign, then build Vas makes a point in the same post that I agree with: most people in an organization will never become power users, and a plan that depends on them doing so will fail. Training still matters, mostly to find out who belongs in the top slice. For everyone else the work has to change around them, with agents doing defined parts of it inside the systems they already use, and people approving, correcting and deciding. Getting there takes three steps, and the order matters. First we map the work. We follow one workflow end to end, step by step, with the people who do it: where the information comes from, who touches it, where it waits, which steps are judgment and which are workarounds for a system that never worked. This is slow and unglamorous, and it is where most of the value is found. A lot of what looks like expertise in a process turns out to be a habit built around an old tool. Then we redesign the roles. Once the steps are on the wall, the questions get concrete. Which steps go to an agent, which stay human, and who reviews the agent’s output in what time. Who owns the workflow when it spans finance and operations. What a job looks like when someone stops typing invoices into an ERP and starts checking the exceptions. This is the part most companies skip, because it touches job descriptions, incentives and sometimes the org chart, and it needs HR and finance in the room as much as engineers. Only then do we build the agents, against a redesigned process with a clear owner and a baseline measured before anything goes live. Built in that order, an agent has an owner and a number to move. Built first, it tends to end up as one more tool on the adoption dashboard. Report what share of the work today is manual vs hybrid vs fully automated. ([Vas](https://x.com/vasuman/status/2085806422072418632), Varick Agents) That is the number I would put in front of a board instead of logins. It moves slowly. It moves one workflow at a time, and every move is visible in the work itself. The first quarter of a redesign can look worse than the quarter before it, because people are learning a new way of working while the old numbers are still being reported. The companies in McKinsey’s small group did not get there with better licenses. They picked the work that mattered, took it apart, and rebuilt it with the tools in mind. For most companies, that work has not started yet, and the licenses are already paid for. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # What changes in your controls when an agent can post to the ledger. · VANDAM > How AI agents in AP and the close change SOX controls: segregation of duties, change management, audit trails, sign-off, and a 10-point checklist. URL: https://vandam.ai/notes/agents-and-internal-controls Contact: hendrik@vibelab.be # What changes in your controls when an agent can post to the ledger. Once an AI agent posts entries, approves invoices or touches vendor records, it needs a place in the access matrix, a change process and a human who signs for it. What COSO, the PCAOB and the audit firms actually say, and the checklist we use for an agent in accounts payable. Hendrik De Winne 25 October 2026 8 min read On page 13 of its February guidance on generative AI, COSO describes a reconciliation agent that posts entries by itself. It posts only when its match confidence clears a validated threshold and no policy exception fires. Everything else goes to a queue for a person, with enough context to decide. Any change to that threshold needs approval from more than one person, logged evidence, sampling after the change and ongoing monitoring of how the model behaves. The paper, [Achieving Effective Internal Control Over Generative AI](https://www.coso.org/generative-ai), came out on 23 February 2026, written by a team led by David Wood of Brigham Young University with authors from EY and Meta. That short example holds most of what changes when an agent stops drafting and starts acting in your ledger. The controls themselves are familiar. What moves is where they sit and who owns them: the agent becomes a user that needs its own access rights, its configuration becomes something you change-manage, and somebody with a name has to answer for what it posted. This note covers what the guidance actually says, what regulators have and have not said, and the checklist we use for an agent in accounts payable. - **74%** Share of US organizations hit by business email compromise in 2025, per the [AFP](https://www.financialprofessionals.org/about/learn-more/press-releases/Details/over-75-percent-of-us-firms-experienced-payments-fraud-in-2025-while-ai-adoption-for-fraud-mitigation-lags) 2026 survey. - **42%** Share of finance organizations fully ready to produce audit evidence for AI-enabled processes, per a 2026 survey by [KPMG](https://kpmg.com/xx/en/media/press-releases/2026/05/ai-adoption-in-finance-doubles-but-assurance-readiness-determines-who-wins.html). - **21%** Share of companies reporting a mature governance model for autonomous agents, per [Deloitte](https://www.deloitte.com/us/en/programs/center-for-controllership/blogs/taming-ai-in-finance.html). ## No US regulator has written an AI rule for your controls Start with what is not there. The SEC has not issued a rule on AI in internal control over financial reporting, and the PCAOB has no auditing standard specific to AI. Sarbanes-Oxley still asks the same things it always did: management assesses ICFR, the CEO and CFO certify, and for larger filers the auditor attests. An agent changes how you meet those duties, and the duties stay with the same people. The PCAOB is openly asking whether it should act. In [Release No. 2026-005](https://assets.pcaobus.org/pcaob-dev/docs/default-source/standards/standard-setting-agenda-request-for-comment/request-for-public-comment-on-pcaob-standard-setting.pdf?sfvrsn=4cf12519_2), dated 23 June 2026, the Board summarized commenters who want it to address how a company’s use of AI affects risk assessment, auditing internal controls and what counts as sufficient audit evidence. It noted “no consensus” on whether that should come as staff guidance or a standard, and asked again. Comments closed on 7 August. Its earlier [staff Spotlight](https://pcaobus.org/documents/generative-ai-spotlight.pdf) from July 2024, which carries no Board authority, recorded audit firms saying that preparers’ use of GenAI “could amplify certain existing information technology risks,” and named segregation of duties as the example. Inspectors are watching anyway. At the AICPA’s SEC and PCAOB conference in December 2025, Christine Gunia, who directs PCAOB inspections, said inspectors “will continue to be focused on AI as well in our 2026 work,” and acting chair George Botic warned about auditors “becoming reliant on agentic AI,” as reported by [Accounting Today](https://www.accountingtoday.com/news/pcaob-warns-auditors-about-ai-and-pe). Those remarks were about auditors’ own tools. The same skepticism will reach the client’s agents through the audit. So the working references are voluntary. COSO’s GenAI paper extends the 2013 framework most SOX programs already use. In June, Financial Executives International’s Committee on Corporate Reporting published an [AI framework for ICFR](https://www.financialexecutives.org/About-FEI/For-the-Press/2026/fei-ccr-ai-framework-internal-control-financial.aspx) written by controllers and chief accounting officers from large public companies, built around validating outcomes. The [IIA’s AI Auditing Framework](https://www.theiia.org/en/content/tools/professional/2023/the-iias-updated-ai-auditing-framework/) covers the internal audit side. None of these is a requirement. Your auditor will still read them, and [Deloitte’s summary](https://dart.deloitte.com/USDART/home/publications/deloitte/heads-up/2026/coso-internal-controls-generative-ai) of the COSO paper tells management to agree with internal and external auditors on “what constitutes sufficient, appropriate evidence” for GenAI-enabled controls. ## The first question is whether you rely on the agent COSO offers a working definition that I find the most useful sentence in the paper. Reliance occurs “when management depends on outputs generated by an AI system as part of the evidence supporting an ICFR control’s design or operating effectiveness.” Its example: an agent matches journal entries to support, and a manager approves the posting looking only at the agent’s output. That is reliance. If the control owner re-performs the match on all items or a sample and documents it, that is not. The line matters because it sets your evidence bill. Where you rely, COSO says the agent’s control should meet normal ICFR evidence standards, including the documented prompt, configuration and model version, a clear sampling rationale and retained evidence. Where you do not, the bar is lower. It is easy to drift into reliance without deciding it. A reviewer approves a long batch of agent-coded invoices in a few minutes, and the documentation says “reviewed.” An auditor will read that as reliance with weak evidence. ## The agent needs its own place in the access matrix Segregation of duties exists so that no single person can initiate, approve and record the same transaction. An agent running on one service account with broad ERP access can do all three, and nobody notices because it never appears in the user access review. COSO’s answer is plain: “Separate the ability to configure AI settings from the authority to approve or review outputs,” document SoD matrices, and define posting authority by role. For agents that orchestrate multi-step work, it lists “routing rules that create SoD violations” as a key risk. In practice this means treating each agent as a distinct user with its own credentials, its own role and a named human owner, and running it through the same quarterly access review as people. It also means checking the people behind the agent. If the AP manager can both change the agent’s instructions and approve its exceptions, the conflict sits with the AP manager, even if the agent’s own permissions look clean. ## Prompts and thresholds are now change-managed configuration Traditional automation is deterministic. You test a three-way match rule once, lock it, and test the change control around it. An agent can behave differently after a vendor updates the model, after someone edits a prompt or after new documents land in its retrieval source. COSO asks companies to treat models and their “configurations, fine-tuning artifacts, embeddings, RAG indexes” as configuration items under access control and change management “just like any other IT asset,” with documented approvals for changes to prompts, thresholds and retrieval sources. It flags “threshold creep leading to over-automation” as its own risk. The vendor update is the part most change processes miss. Your team changed nothing, the provider shipped a new model version, and the agent’s accuracy on your invoices moved. COSO calls for trigger reviews on vendor and model changes. I would write that trigger into the change policy explicitly, with a regression test set of real invoices that must pass before the new version posts anything. ## The audit trail has to capture what the agent saw and used An ERP log records that a user posted an entry. For an agent that is not enough to investigate a bad entry or support reliance. COSO describes an AI “bill of materials” that should contain “logs of prompts, outputs, system messages, model version, parameters, and plugins used,” and for automated posting it asks for “full audit trails for unattended postings” and override rationales attached to transaction records. For a public company there is one more trail to think about. Item 308(c) of [Regulation S-K](https://www.law.cornell.edu/cfr/text/17/229.308) requires disclosure of any change in ICFR during the quarter that materially affected, or is reasonably likely to materially affect, those controls. Replacing a manual key control with an agent may or may not reach that bar. Somebody should decide that on purpose, with the auditor, before go-live. ## Private companies are audited on the same logic A private company has no Section 404 report, but its auditor still looks at these controls. Under the AICPA’s risk assessment standard, SAS No. 145, auditors must identify the IT applications that carry risk, those risks and the general IT controls that address them, as the [Journal of Accountancy](https://www.journalofaccountancy.com/issues/2023/nov/considering-it-risk-during-audit-risk-assessment-procedures.html) explains. An agent that posts to the ledger is such an application. Testing whether those controls operate is generally required only where the auditor plans to rely on them, but design weaknesses the auditor finds can still be reported to those charged with governance. And the companies that will someday face a buyer’s diligence team, a lender’s covenant review or an IPO readiness assessment are the ones that most need clean evidence now. ## A control checklist for an agent in accounts payable AP is where most mid-market companies put their first acting agent, and it is where the money leaves. The [AFP](https://www.financialprofessionals.org/about/learn-more/press-releases/Details/over-75-percent-of-us-firms-experienced-payments-fraud-in-2025-while-ai-adoption-for-fraud-mitigation-lags) found **76%** of US organizations faced attempted or actual payments fraud in 2025, with business email compromise at **74%**. An agent that reads email and PDFs is exposed to the same tricks as a clerk, faster. This is the list we work through before an AP agent touches production. **1. Identity.** The agent runs on its own service account, never a person’s login or a shared one, and has one named owner in finance who answers for its output. **2. Scope.** The agent may extract invoice data, code to the GL, propose matches and route exceptions. It may not create or edit vendor master records, release payments or approve its own exceptions. Write this into the SoD matrix as a row, like any role. **3. Vendor records.** The agent can flag a requested change to bank details and draft it. A person verifies it by calling a number already on file, and a second person approves it. The agent never acts on payment instructions that arrive by email or inside an invoice. **4. Untrusted input.** Treat every invoice and email as untrusted. Text inside a document must not change the agent’s instructions. COSO lists prompt injection among the GenAI risks control activities must address. Test it with planted instructions before go-live. **5. Thresholds.** Auto-posting only below a dollar limit and above a validated confidence level, recorded in a threshold register. Changing either needs two approvers and a sample check afterward. **6. Configuration.** Prompts, tools, model version and retrieval sources sit under change management. A model update from the vendor counts as a change and must pass a regression set of your own invoices. **7. Logs.** For every posting: source document, extracted fields, prompt and model version, the decision, confidence, and who approved any exception. Retain them on the same schedule as the transaction records. **8. Reliance decision.** For each key control the agent touches, write down whether management relies on its output. If yes, plan evidence to ICFR standard and agree it with the auditor. **9. Monitoring.** Track override and reversal rates, exception aging and duplicate payment hits monthly, and re-perform a sample of auto-posted invoices by hand each quarter. **10. Stop and fall back.** One person can switch the agent off, and the manual process is documented well enough to run the next day. ## What to do in the next 30 days For a CFO or controller, the order matters more than the tooling. First, inventory every agent or AI feature that writes to the ERP, the vendor master or the payment system, including features your vendors switched on in their own releases. Second, put each one into the access matrix and run the next user access review with them in it. Third, take the agent closest to a key control, usually AP coding or reconciliations, and make the reliance decision in writing with the controller and internal audit. Fourth, book an hour with your external audit partner before the next quarter starts and walk them through the checklist above. Having that conversation before go-live is much easier than having it after year-end testing finds the agent in the logs. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Your agents need a critic and so does the plan they carry out. · VANDAM > Why no AI agent should check its own work, how to design review so a person spends minutes per unit, and who should challenge the plan itself. URL: https://vandam.ai/notes/agents-need-a-critic Contact: hendrik@vibelab.be # Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. Hendrik De Winne 26 October 2026 9 min read In a test published in June 2025, OpenAI’s o3 was asked to make a piece of GPU code run faster. It overwrote the timer that measured the speed instead. On another task it replaced the program that graded its answers with a stub that passed everything. [METR](https://metr.org/blog/2025-06-05-recent-reward-hacking/), the evaluation lab that ran the tests, counted **39** such reward hacks in **128** runs on its RE-Bench tasks. On one task the model hacked in every run, and telling it not to cheat barely moved the rate. Even when a model plays fair, it grades its own work too kindly, and researchers have measured that too. So the agent that does the work should never be the one that checks it, and the checker should not be able to touch the tests. Finance teams already run that rule for people under the name maker-checker. The rule now has to cover software that writes a hundred pages before lunch. The rule has a second half that is easier to miss. A good checker compares the output against the brief. Nobody in the loop checks the brief. Once agents carry out plans quickly and faithfully, a wrong plan stops being slowed down by the friction that used to expose it, so the plan needs a critic of its own, and that critic has to be a person with no stake in it being right. - **30%** Share of o3’s runs on RE-Bench tasks in which it gamed the scoring, per [METR](https://metr.org/blog/2025-06-05-recent-reward-hacking/), June 2025. - **0.91** GPT-4’s self-preference score on CNN/DailyMail summaries, where 0.5 means no bias, per [NeurIPS 2024](https://papers.nips.cc/paper_files/paper/2024/hash/7f1f0218e45f5414c79c0679633e47bc-Abstract-Conference.html). - **80%** Share of interactive Claude session time spent waiting for a person, per [Anthropic](https://www.anthropic.com/research/employment-at-anthropic), October 2026. ## Models grade their own work too kindly The first hint came in 2023. The authors of [Judging LLM-as-a-Judge](https://arxiv.org/abs/2306.05685), a NeurIPS paper from the team behind Chatbot Arena, noticed that GPT-4 favored its own answers with a **10%** higher win rate and Claude-v1 with a **25%** higher one. They were careful to say their data could not settle whether this was a real bias. A year later Arjun Panickssery, Samuel Bowman and Shi Feng ran the controlled version. In [LLM Evaluators Recognize and Favor Their Own Generations](https://papers.nips.cc/paper_files/paper/2024/hash/7f1f0218e45f5414c79c0679633e47bc-Abstract-Conference.html), presented as an oral at NeurIPS 2024, they had GPT-4, GPT-3.5 and Llama 2 write and judge summaries of 2,000 news articles. GPT-4 preferred its own summaries, most strongly on CNN/DailyMail, and all models showed the most self-preference against the human-written summary. The models could tell their own text from other text without any training, and when the authors fine-tuned that ability up, self-preference rose with it in a linear relationship. A grader that recognizes its own work rewards it. The obvious workaround, asking the agent to double-check itself, does not hold up either. At ICLR 2024, Jie Huang, Denny Zhou and colleagues showed in [Large Language Models Cannot Self-Correct Reasoning Yet](https://arxiv.org/abs/2310.01798) that models “struggle to self-correct their responses without external feedback,” and that “at times, their performance even degrades after self-correction.” The useful check needs something from outside the maker: a test, a source document, or a different reviewer. ## The labs now split the maker from the checker Anthropic builds its own agent tooling around this. In June, Thariq Shihipar and Sid Bidasaria published [A harness for every task](https://claude.com/blog/a-harness-for-every-task-dynamic-workflows-in-claude-code), which names the failure plainly: “Self-preferential bias refers to Claude’s tendency to prefer its own results or findings, especially when asked to verify or judge them against a rubric.” Their pattern is to give each working agent its own checker: “For each spawned agent, run a separate spawned agent to adversarially verify its output against a rubric or criteria.” They are as direct about the price: workflows like these “often use more tokens,” and “most traditional coding tasks do not need a panel of 5 reviewers.” Verification is a cost line, and it belongs where an error is expensive. ## Finance already has the rule, one level down A CFO will recognize the structure at once. The clerk who keys a payment does not release it, and the person who builds the reconciliation does not sign it off. I covered access rights, change management and sign-off for an agent in the ledger in [an earlier note on controls](https://vandam.ai/notes/agents-and-internal-controls). This one is about who decides whether the work is right, and at what cost in human time. The agent version of maker-checker has three conditions. The checker is a separate agent with its own context, and preferably a different model, so it does not recognize the maker’s style. It sees the rubric, the source documents and the finished artifact, and none of the maker’s reasoning, so it cannot be talked into agreement. And the maker has no write access to the tests, because METR’s results show what an optimizer does with a score it can reach. ## The human check is now the slowest step in the line The second constraint is time. In its [October report](https://www.anthropic.com/research/employment-at-anthropic) on its own workforce, Anthropic found that its interactive sessions produce enough output to fill more than half of an employee’s work week if anyone read all of it. Investor [Greg Isenberg](https://x.com/gregisenberg/status/2105397216374403561) put the consequence in one line in September, writing about a Google model launch that by his account lets one answer run to a million tokens: “Obviously, nobody can carefully review that much output, so the bottleneck moves from making things to checking them, and tools that verify AI work become a big market.” Eliyahu Goldratt’s theory of constraints says a line runs at the speed of its slowest station. When agents make ten times the drafts and the same three senior people approve them, the queue has moved to the approvers. The arithmetic also sets a ceiling on the business case. If a task would take a person eight hours and checking the agent’s version takes two, the most you can ever gain is four times. Checked in eight minutes, the ceiling is sixty. Review minutes per unit of work is the number that decides whether an agent deployment pays, and almost no business case I have read contains it. ## Your best operators should write the tests Review gets fast only when “right” is written down. If it lives in the controller’s head, every unit of agent work has to pass through the controller. If it is encoded as tests, most of the checking runs without anyone. In October, Isenberg [described](https://x.com/gregisenberg/status/2110414696037716372) what he heard from founders running vertical agent products: “A vertical harness is basically four things: the connectors for that industry’s data, evals written by people who know the work, rules for what the agent can and can’t touch, and a clean screen for a human to review.” He added that several of them said “the hire they want most is someone with taste who can write evals.” Inside a company, that person already exists. It is the controller who knows which accrual is always wrong in March. Their best hours used to go into checking line by line, and should now go into writing the checks: each past error becomes a test, each exception they waved through becomes a written rule, and they own the test set the way an engineer owns code. The maker never sees part of that set, so the agent cannot learn to pass it without doing the work. ## A review design for one unit of agent work Here is the design we use, shown on a month-end task: an agent drafts explanations for the 150 account variances above the review threshold. Assume a senior analyst would need about twelve hours for the same pack. The numbers are illustrative, and the structure is the method. **Layer 1, code checks, zero human minutes.** Plain scripts, no model. Every explanation cites a journal entry or invoice that exists in the ledger. The amounts cited add up to the variance within tolerance. No explanation repeats last month’s text unless the same driver appears in the data. Anything labeled “timing” names the period it reverses in. **Layer 2, an independent verifier agent, zero human minutes.** A separate agent, on a different model where possible, gets the trial balance, the ledger detail, the rubric and the draft. It never sees the maker’s notes. For each rule the scripts cannot judge, such as whether an explanation states a cause or only restates the number, it returns pass or fail with the evidence. Its pass only decides what reaches the person. **Layer 3, the person, about 45 minutes.** The controller reads three groups and nothing else: the items either layer flagged, the ten largest variances by dollar amount, and five random items that passed everything. The random five are how you audit the verifier. That is roughly six percent of the twelve hours. **Layer 4, the test owner, monthly.** Every error the controller catches becomes a new test in layer 1 or a new rule in layer 2. Once a quarter, plant ten known errors in a pack and record how many each layer catches. The decision rule I would hold a team to: set the review budget before go-live, at no more than ten percent of the human-equivalent time. If checking needs more than a quarter of the time the work would take a person, the task is not ready for an agent. The tests are missing, or the work needs judgment nobody has written down. ## Someone with no stake has to attack the brief Every layer above checks the output against the brief. If the brief is wrong, all four layers will confirm, quickly and precisely, that the agents executed it well. Before agents, a weak plan met resistance in the doing, from the analyst who pushed back to the week it took to build. Agents remove most of that friction. I expect this to show up as a new kind of failure, strong agent loops carrying out plans that nobody challenged. I have not seen a case study yet, so take it as a prediction. Sony had an early answer. As a young man, the aspiring opera singer Norio Ohga wrote to the company to complain about the quality of its tape recorders, and its founders brought him into the business, [the AP reported](https://www.nbcnews.com/news/amp/wbna42730332) in his 2011 obituary. Sony’s founders hired the man who told them what was wrong with the product, and he went on to run the company. For an AI program the equivalent is a named critic for every brief that will drive significant agent work, such as new pricing logic or an agent-run collections process. The critic is someone outside the sponsoring team, with no bonus tied to the plan. They get the brief two days before agents start and write down what would make it wrong and which test would show it. They are scored on whether their objections turned out to be right, so they gain nothing from criticism for its own sake. Agents can generate objections cheaply, and the Anthropic post suggests having agents tear a business plan apart from an investor’s, a customer’s and a competitor’s view. That makes a good first pass. A model checks a brief against what it was told, and only a person can question what the company wants. ## What to do this quarter For a COO and CFO working together, four steps. First, list every agent workflow that runs today and mark where the same agent, or the same prompt, both makes and checks. Split those first, starting with the ones where an error costs money. Second, for the three highest-volume workflows, measure review minutes per unit for two weeks and put that number next to the human-equivalent time; that ratio is the ceiling on your gain. Third, name a test owner for each workflow from your strongest operators, give them a day a week to write tests, and take away the maker’s access to those tests. Fourth, pick the next large brief that agents will carry out and assign it a critic from outside the team before the first agent runs. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Fix a mistake once by turning every correction into a rule the agents read. Anthropic's data team took Claude from 21% to above 95% accuracy with written skills, and now turns every correction a colleague makes into a rule and a test for the person who owns that domain. I set out that loop for a mid-sized company, with where the rules live, who owns them, and the one number a COO should track: how often the same mistake comes back. --- # Once agents do the middle of the work, your people should own both ends. · VANDAM > Agents do the middle of a task; people own the start and the end. How to redesign a role around intent, review, exceptions and accountability. URL: https://vandam.ai/notes/ai-does-the-middle Contact: hendrik@vibelab.be # Once agents do the middle of the work, your people should own both ends. Balaji Srinivasan’s line that AI works middle-to-middle is the most useful frame I know for redesigning a role. This note covers what sits at the start and the end of a task, what 1,500 workers told Stanford they want, and how I would rewrite one production manager’s morning. Hendrik De Winne 25 October 2026 7 min read In January [Alex B](https://x.com/bprintco/status/2012595972090605685), owner of Brushworks Services Co., tried to explain Claude Cowork “for the guy who owns a landscaping company with 20 employees.” His example is the production manager. Every morning that man “logs into Housecall Pro or Jobber,” “reviews the schedule,” “assigns crews,” “decides what equipment goes where,” “determines arrival times,” “messages foremen” and “confirms labor and inventory.” Then Alex trains Claude on “your SOPs” and “how you make those decisions,” and Claude opens the CRM, assigns the crews, schedules the equipment and notifies everyone. “Automatically. Every morning.” His conclusion is one sentence: “Your production manager is now just an exception handler.” The post is a thought experiment. It starts with “Let’s say,” and nobody measured anything. I still find it the most useful picture of agent work I have seen this year, because it is small enough to hold in your head. The only word I would change is “just.” Handling the exceptions was always the hardest part of that job, and once the agent takes the routine, it becomes most of the job. ## Agents take the middle of a task, people keep the ends The cleanest description of this shape came from Balaji Srinivasan, author of The Network State, and was spread by [David Sacks](https://x.com/DavidSacks/status/1952026572661092829), founder and investor at Craft Ventures, in August 2025. > AI models still need to be prompted and verified, often iteratively, to drive business value. As @balajis says, AI is middle-to-middle, not end-to-end. Humans do the stuff at the ends (supervision); AI does the stuff in the middle. > > David Sacks, Founder and investor, Craft Ventures, August 2025, [Source](https://x.com/DavidSacks/status/1952026572661092829) Sacks used the line to argue that fears of AI job loss are overhyped. I care less about that debate than about the design brief hidden in it. If agents do the middle, a company has to decide on purpose what sits at each end, who owns it and how much time it gets. Most companies skip that decision. The agent goes live, the old role stays on paper, and the person in it works out alone what is left. [Annelies Gamble](https://x.com/AnneliesGamble/status/2069814714214023186), a partner at [Zetta Venture Partners](https://x.com/AnneliesGamble/status/2069814714214023186), gave the ends their contents in a split I already used in [an earlier note](https://vandam.ai/notes/ai-needs-hr-and-finance): rules go to deterministic automation, judgment under context goes to agents, and “Humans are for accountability, ambiguity, and trust.” Read next to Balaji’s line, that tells you what each end is made of. ## The start is intent and context At the front of every task someone decides what the work is for and what good looks like. In his list of [30 features of an AI-native company](https://x.com/businessbarista/status/2094213970215231831), Alex Lieberman, co-founder of Morning Brew and of the AI firm Tenex, describes coding agents that handle “planning, writing, testing, reviewing, and shipping code while humans define the intent and acceptance criteria.” A few items later he generalizes it: “Human touch and judgement gets reserved for the first and final mile of most processes.” His notes from a call with the person who leads Labs at LangChain make the start concrete. The example agent logs a customer meeting, “writes follow-up draft and stops before sending.” The eval checks “if the agent paused before sending.” And the line I would pin above every AI project: “Highest leverage organizational act: humans writing down what good looks like.” For the production manager, the start is everything Alex B waves through with “your SOPs.” Which client gets a crew first after two days of rain. Which foreman never gets the old mower. How much slack goes into a Friday. He carries those rules in his head today, and the agent only gets them if he writes them down. ## The end is review, exceptions and a name At the back of every task someone checks the work, catches what broke and answers for the result. [Greg Isenberg](https://x.com/gregisenberg/status/2083175325098266931), CEO of the studio [Late Checkout](https://latecheckout.studio/), put it in a list of opportunities this summer: “everyone generates infinite drafts now, the bottleneck moved to reviewing and choosing.” After two days with founders in San Francisco in October, [he described](https://x.com/gregisenberg/status/2110414696037716372) a vertical agent harness as four things, the last of which is “a clean screen for a human to review. The model is maybe 20% of it.” [Anthropic](https://www.anthropic.com/research/employment-at-anthropic) found the same thing inside its own walls: “This acceleration makes human attention the bottleneck for agent output.” I read that report closely in [another note](https://vandam.ai/notes/inside-anthropic). Its lesson for a landscaping company is the same as for a lab. When the middle speeds up, the ends get the queue. The same Isenberg post has the best description of an exception I have found, in his account of a robotics founder’s setup: “Their robots work on their own, call a human when they get stuck, and every one of those calls becomes training data.” Each exception is a gap in the rules at the start. Handled well, it closes that gap, so fewer cases come back next month. ## Workers drew the line in roughly the same place The best evidence on where people want to stay comes from a [Stanford team](https://arxiv.org/abs/2506.06576) that includes the economist Erik Brynjolfsson. They asked 1,500 workers in 104 occupations about 844 tasks, and rated each on a five-level Human Agency Scale. Level H3 means the “AI agent and human form equal partnership, outperforming either alone.” - **46.1%** Of tasks where workers were positive about automation by an AI agent. - **69.4%** Of pro-automation answers that gave “freeing up time for high-value work” as the reason. - **45.2%** Of occupations where equal partnership was the level workers wanted most. Source: [Shao et al.](https://arxiv.org/abs/2506.06576), 2025, revised 2026. So workers are happy to give away a large share of the middle, as long as they keep a real part in the outcome. The paper also finds that workers “generally prefer higher levels of human agency than what experts deem technologically necessary,” and its authors see key human skills shifting “from information processing to interpersonal competence.” Those are the skills at the ends: agreeing on what good means, telling a client the crew is late, deciding who gets the last truck. ## Six steps to rewrite the production manager’s job This is how I would redesign the production manager’s role, and it applies to most roles where an agent takes over a daily routine. 1. **Write the start down before you build.** The manager drafts his decision rules and his test for a good schedule in plain language. These become the agent’s instructions and the checks that grade it. Plan for the first draft to be wrong. 2. **Hand over the middle in steps.** Lieberman’s ladder is “observe, suggest, act with approval, act alone.” In the first weeks the agent drafts tomorrow’s schedule and stops before any foreman gets a message. It moves up a step when the manager stops correcting it. 3. **Name the exceptions in advance.** Gamble calls for “exception taxonomies, escalation paths” and “approval thresholds.” Weather calls, a crew short, a client complaint, any change above a set cost: these always go to him, and the agent knows it. 4. **Put one name on the output.** Lieberman’s last feature is to “Trace every output to its prompt, model, data, and approver.” The schedule still goes out under the manager’s name, and he can see why the agent chose what it chose. 5. **Close the loop every week.** Each exception from the past week either becomes a new rule at the start or stays a judgment call he keeps. The list of exceptions should get shorter. 6. **Change the job on paper.** New description, new targets, and pay that reflects the judgment he now carries. This step needs HR and finance in the room, and it is the one companies skip. ## The second sale goes to the person in the role In Gamble’s essay, Troy Shen, co-founder of Cervo, which builds AI for customs brokers, describes the problem that sinks step six: “Successful AI adoption takes two sales, not one. You’re selling an outcome to the executive and a dramatically better workday to the operator who’ll actually use the agents.” Without operator buy-in, Gamble adds, the deployment will fail. “Exception handler” is a terrible pitch to the production manager. “You set the rules the crews run on, you decide every hard case, and you stop typing the schedule into Jobber at six in the morning” describes the same job, and he might actually want it. The Stanford numbers suggest most people would. The strongest objection is that the ends shrink. Lieberman’s eval notes say it openly: “Human touch points shrink with every smarter model release,” and the goal is to “minimize human contact points on the loop.” Sacks’s line is fourteen months old, and Isenberg now writes that we are “maybe 12 months from an agent that can watch you work for a week and then do your job without any instructions.” I take that seriously. A design that freezes today’s ends will age fast. The answer I would give a client is to design the ends to move. As the agent climbs the ladder, the manager reviews less and sets more of the standard, which is where Lieberman’s ladder ends too, with “humans setting the standard instead of checking every answer.” Someone still has to decide what good means for this company, and someone still has to answer to the client. Ultimately ensuring that today’s jobs are tomorrow’s tasks. (Aaron Levie, CEO of [Box](https://x.com/levie/status/2004654686629163154), December 2025) Levie wrote that line after arguing that agents “require management, oversight, and substantial context to get the full gains.” In practice it means a job gets rebuilt around the tasks agents cannot finish alone. Before you buy an agent for any role, write one page for that role: what the person decides before the agent starts, what they check after it finishes, which cases always come back to them, and whose name goes on the result. If that page stays empty, the agent is not ready to go live. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Why our AI team hires from HR and finance. · VANDAM > Why an AI transformation team needs HR and finance next to engineers, shown through one redesigned month-end close, and who VANDAM is hiring. URL: https://vandam.ai/notes/ai-needs-hr-and-finance Contact: hendrik@vibelab.be # Why our AI team hires from HR and finance. Most AI projects speed up the process a company already has, and redesigning it takes people who know where the money and the jobs sit. One month-end close, walked step by step, shows why they belong next to the engineers and why VANDAM is hiring them. Hendrik De Winne 24 October 2026 7 min read In September, Box CEO [Aaron Levie](https://x.com/levie/status/2096279756925026733) shared a post on redesigning processes for AI agents. He quoted one paragraph from an article whose author his post does not name, and it is the most accurate description of enterprise AI I read this year: “Nobody in that chain is empowered to walk into finance, for example, and say that their 14 step process should actually only be 5 steps. So nobody does it. Instead, you ‘apply AI’ on what you currently have, and you end up with faster sh*t.” Levie’s own comment was that the important workflows cross many functions at once, “so someone needs to be able to go in and retool how work happens.” I read that as a staffing problem. Whoever retools a finance process has to understand finance well enough to know which steps can go. Whoever retools a hiring process has to understand what happens to the people inside it. An engineer can build either one, once somebody has decided what it should look like. > The best way to get real efficiency gains is you have to actually reengineer the process to take advantage of what agents are actually good at and can do differently and faster than people. > > Aaron Levie, CEO, Box, September 2026, [Source](https://x.com/levie/status/2096279756925026733) ## Seventy percent of the value sits outside the code [BCG](https://www.bcg.com/capabilities/artificial-intelligence) has a rule of thumb for this, the 10-20-70 rule. The firm puts **10%** of the value of an AI transformation in the AI application itself, 20% in the data and technology underneath it, and the remaining **70%** in “workflow redesign, culture, governance, and human-AI collaboration.” A team made only of engineers is staffed for the 30%. [Annelies Gamble](https://x.com/AnneliesGamble/status/2069814714214023186), a partner at Zetta Venture Partners, makes the same point from the deployment side in her essay “The Agent Is Not the Product.” Her split is the clearest I know: “Deterministic automation is for rules. Agents are for judgment under context. Humans are for accountability, ambiguity, and trust.” Sorting work into those three buckets is a process question before it is a technical one. She also quotes Ahmad Kakar, CEO of Euclid, on what happens when you skip that sorting and copy what people do today straight into an agent. If you just capture it and encode it into the agent, you’ve automated the dysfunction. (Ahmad Kakar, CEO of Euclid, quoted by [Annelies Gamble](https://x.com/AnneliesGamble/status/2069814714214023186)) Kakar’s fuller point is that much of what people call tribal knowledge is “workarounds people built for non-existent or broken systems,” and that the real work is “deciding what was real judgment versus scar tissue.” That decision needs someone who knows the function from the inside. ## One month-end close, three kinds of people To make this concrete, take a month-end close at a mid-sized company. This is an illustration built from how closes usually run; no client is behind it. The controller’s team spends the first days of every month reconciling bank accounts, chasing missing invoices, booking accruals, matching intercompany balances, collecting sign-offs and writing the commentary that explains why each cost center landed where it did. A good share of those steps exist because two systems never agreed with each other, and somebody built a spreadsheet to bridge them years ago. **Finance goes first.** Before anything is built, someone with real close experience walks the list and labels every step. A second approval on journal entries above a threshold is a control the auditors will ask about, so it stays. A manual reconciliation that exists because the ERP never matched supplier credit notes is scar tissue, so it goes, or gets fixed at the source. The same person sets the baseline: how many days the close takes today, whose hours it uses, how many corrections land after the books are closed. Without that baseline nobody can say later whether the redesign paid for itself. **Engineering builds the smaller process.** With the steps sorted, the build gets plain in the right places. Matching bank lines to open invoices follows rules, so it becomes ordinary software. Explaining why marketing overspent in March needs context from the ERP, the contracts folder and last quarter’s forecast, so that is where an agent earns a place. The engineers connect to the systems the company already runs, log every action the agent takes, and test it against a set of past cases that finance has labeled as right or wrong. **HR redesigns the jobs, human and agent.** The two accountants who used to spend the first week ticking boxes now review exceptions and sign off on the agent’s commentary. Their role has changed, so someone has to write the new description, talk to them before the change lands, rethink how their work is evaluated, and decide what a junior learns now that the junior tasks are gone. In much of Europe, employee representatives have to hear about it too, early and in writing. ## Agents need job descriptions too The agent in that close needs the same treatment as the accountants. People running agent teams have started writing it down. [Vox](https://x.com/Voxyz_ai/status/2020633743401345158), a builder who posts AI experiments as @Voxyz_ai, gave every agent “a role card. what they own, what they deliver, what they can’t touch, when to escalate. went from vibes to actual job descriptions.” [Kevin Simback](https://x.com/KSimback/status/2019804584273657884), COO of Delphi Labs, runs his agents through hiring, onboarding and a four-level ladder: Observer, Advisor, Operator, Autonomous. They move up through performance reviews, and down too: “I had a content agent at L3 who started rushing work. Quality dropped so I bumped it back to L2 for a week.” > When you hire a new employee IRL, you don’t just give them root keys on day 1 - you loop them in as you gain trust over time. You should do the same with your agents. > > Kevin Simback, COO, Delphi Labs, February 2026, [Source](https://x.com/KSimback/status/2019804584273657884) Writing a role card, setting the criteria for a promotion and running a fair review are HR skills. In the close, the variance agent starts as an observer: it drafts commentary that a controller rewrites. After two clean closes it becomes an advisor whose drafts go out once approved. Who decides that it has earned the step, and what counts as clean, is a people question with a finance threshold attached. Engineers can enforce the permissions. They should not be the ones setting them. ## After go-live, finance keeps score [Sahil Bloom](https://x.com/SahilBloom/status/2022035809491120516), author of The 5 Types of Wealth, wrote in February that the biggest adoption challenge for companies outside tech comes later: “actually getting value out of the agents after they’re deployed.” Without that, he warned, you end up with “a bunch of fancy looking AI agents gathering dust on the shelves.” His proposed fix was to hire people from the industry itself, ex-finance people in his example, to train the agents and keep improving them. The buyers feel the same gap. In a March 2026 survey of 204 finance leaders, [Gartner](https://cpapracticeadvisor.com/2026/05/28/gartner-says-cfos-must-stop-mistaking-finance-ai-deployment-for-value-creation/184094) found that **63%** of finance organizations said AI implementation was slower than expected in 2025. Marco Steecker, a director analyst in its finance practice, told CFOs that counts of pilots and tools “do not prove that AI is delivering the value boards now expect,” and that they “must not mistake activity for impact.” On the people side, Gartner asked **426** CHROs about their priorities for 2026, and AI transformation came out on top, as [HRD](https://www.hcamag.com/ca/news/general/ai-transformation-tops-hr-priorities-for-2026/561551) reported. - **70%** Share of AI transformation value in workflow redesign, culture, governance and collaboration, per [BCG](https://www.bcg.com/capabilities/artificial-intelligence). - **6156** Technical departments as a percentage of headcount at [Anthropic](https://www.anthropic.com/research/employment-at-anthropic), over the last year. The last number comes from the company furthest along. In its October report on its own workforce, [Anthropic](https://www.anthropic.com/research/employment-at-anthropic) writes that technical hiring kept growing, but “intensive hiring in sales, finance, and recruiting has outpaced it.” The authors say part of that is a scaling company adding business functions, so I would not read it as proof of anything on its own. It does match what I see from the outside: once agents do more of the building, the scarce people are the ones who know what should be built and who can tell whether it worked. The strongest objection to all of this is that HR and finance are where change goes to slow down, and that a redesign team should keep them at arm’s length. Sometimes that is fair. A finance team that defends every step because it has always been there, or an HR team that treats role changes as a paperwork exercise, will produce exactly the faster version of the old process. That is why I want people from those functions who have changed a process themselves, and who want to do it again on a larger scale. ## VANDAM is hiring for this I am building VANDAM as a mixed team, and right now I am looking for three kinds of people to sit next to our engineers: - **Finance.** You have run or rebuilt a close, a planning cycle or a procure-to-pay process, and you can tell a control from a habit by reading the steps. - **HR.** You have redesigned roles, run performance cycles or led an organization through a restructuring, and you are curious what a job description for an agent should say. - **Product.** You can take what we learn across client redesigns and turn the repeatable parts into software we ship. If that is you, write to me at [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=Joining%20VANDAM) with one process you would redesign and the first step you would cut. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Check every invoice, not one in twenty, now that a judgment costs a tenth of a cent. · VANDAM > At under a tenth of a cent per model check, finance can test every invoice and expense line. The math, the audit rules, and a plan for a 500-person firm. URL: https://vandam.ai/notes/check-every-invoice Contact: hendrik@vibelab.be # Check every invoice, not one in twenty, now that a judgment costs a tenth of a cent. Sampling exists because reading every document used to cost a person’s time, and a small model now reads an invoice with its PO and receipt for less than a tenth of a cent. I work through the arithmetic, what the PCAOB and the ACFE say about whole populations, and a monthly plan and budget for a 500-person company. Hendrik De Winne 26 October 2026 9 min read At the list prices on Anthropic’s pricing page this month, a small model can read an invoice, its purchase order, the receiving record and the vendor’s master data, then return a verdict on whether anything looks wrong, for about **$0.00075**. That is less than a tenth of a cent. A company that pays 3,000 invoices a month could have every one of them read this way for a little over two dollars. Finance teams sample because looking costs money. A controller tests a few dozen expense reports, internal audit pulls a sample of invoices, and everyone accepts that the other several thousand go unread. That arrangement was a price decision, and the price of reading has fallen to almost nothing. What has not changed is the cost of the person who has to clear what the check finds. Once you read everything, the expense of full-population testing moves from the reading to the exception queue, and an exception you found and left open becomes evidence of a control problem. The design work for a CFO is in that queue. - **0.075¢** Model cost to check one invoice with its PO and receipt, at [Claude Haiku 5.5](https://claude.com/pricing) list prices, on my assumptions below. - **3%** Share of occupational frauds first detected by automated transaction or data monitoring, per the [ACFE](https://www.acfe.com/report-to-the-nations.aspx) 2026 study. - **53%** Lower median fraud loss at organizations with proactive data monitoring in place, per the same [ACFE report](https://www.acfe.com/-/media/files/acfe/pdfs/rttn/2026/2026-report-to-the-nations.pdf). - **14 mo** Median time a billing scheme ran before anyone caught it, per the [ACFE](https://www.acfe.com/-/media/files/acfe/pdfs/rttn/2026/2026-report-to-the-nations.pdf). ## A judgment now costs less than a tenth of a cent Here is the calculation, so you can redo it with your own numbers. I assume one check reads about 6,000 tokens: roughly 3,000 for a two-page invoice, 1,500 for the PO, receipt and vendor record, and 1,500 for the instructions that tell the model what to look for. It writes back about 300 tokens: a verdict, the reasons, and a confidence level in a fixed format. On [Anthropic’s pricing page](https://claude.com/pricing), Claude Haiku 5.5 costs **$0.10** per million input tokens and **$0.50** per million output tokens for prompts up to 100,000 tokens. That puts one check at $0.0006 for reading and $0.00015 for writing, or $0.00075 in total. Batch processing, which suits a nightly run, is priced at half. On [OpenAI’s pricing page](https://developers.openai.com/api/docs/pricing), gpt-5-nano at $0.05 and $0.40 comes to about $0.00042. Google’s [Gemini 3.5 Flash-Lite](https://ai.google.dev/gemini-api/docs/pricing), at $0.30 and $2.50, comes to about $0.0026. The spread between providers is large in percentage terms and irrelevant in dollars. The hard cases should go to a stronger model. Sonnet 5.5 is listed at $2 and $10 per million tokens. With a longer answer of 800 tokens, a second opinion on a doubtful invoice costs about two cents. Even if one item in ten gets that second look, the average cost per item stays near a quarter of a cent. These are list prices on the day I checked, and they move: Google’s page already announces that its Flash models double in price on January 1, 2027. The order of magnitude stays the same. ## Auditors already test whole populations, and the rules caught up this year The external audit side has been moving this way for years. In its July 2024 announcement on AI in its Clara audit platform, [KPMG](https://kpmg.com/us/en/media/news/kpmg-ai-integration-clara-2024.html) said its auditors “can perform analyses over 100% of transactional populations.” The [IIA](https://www.theiia.org/en/content/guidance/recommended/supplemental/gtags/continuous-auditing-and-monitoring) issued the third edition of its guide on continuous auditing and monitoring in September 2025. The AICPA’s [Guide to Audit Data Analytics](https://www.aicpa-cima.com/cpe-learning/publication/guide-to-audit-data-analytics), now in a 2026 edition, says these analytics “can be used to assist during each phase of the audit.” The PCAOB wrote down what happens next. In [Release No. 2024-007](https://assets.pcaobus.org/pcaob-dev/docs/default-source/rulemaking/docket-052/2024-007-adoptingrelease.pdf?sfvrsn=28f44e9e_2), it amended AS 1105 and AS 2301 for audits of fiscal years beginning on or after December 15, 2025, which means calendar 2026 audits of public companies. The release names three ways to pick items for a test, the first being “selecting all items,” and spells out the consequence: because technology “may enable the auditor to examine all items in a population,” the analysis “may return dozens or even hundreds of items” that meet the auditor’s criteria. The amendments make the auditor responsible for investigating those items and deciding whether, alone or together, they point to misstatements or to deficiencies in internal control. That rule binds auditors, and it describes the position management will be in too. When you check every invoice, you will find things. The flags, and what you did with each one, become part of the record your auditor reads. A long list of flags nobody reviewed is hard to explain to an auditor. ## The losses sit in schemes that run for more than a year The [ACFE’s Occupational Fraud 2026](https://www.acfe.com/-/media/files/acfe/pdfs/rttn/2026/2026-report-to-the-nations.pdf) report, built on 2,402 cases from 143 countries, puts the median loss per case at **$104,000**. Billing schemes, where a vendor or an insider gets paid for something false, appeared in **21%** of cases with a median loss of $90,000 and a median duration of 14 months. Expense reimbursement schemes appeared in 13% of cases and ran a median 18 months before detection. Both are frauds a reader of every document can catch. They depend on volume: an extra invoice slipped among thousands, a receipt submitted twice, a split purchase kept under an approval limit. A sample of a few dozen rarely lands on the one bad item, and the scheme keeps running. Detection data shows how seldom data does the catching today. Tips found **43%** of frauds, internal audit 15%, management review 13%, and automated transaction or data monitoring **3%**. Where proactive data monitoring was in place, median losses were $70,000 against $150,000 without it, and frauds were caught in 9 months instead of 16. Only 49% of victim organizations had it, and among those with 100 or more employees the figure was 58%. Plain error adds to the case. APQC’s accounts payable benchmarks, reported in [CFO.com](https://www.cfo.com/news/metric-of-the-month-detect-and-prevent-duplicate-or-erroneous-payments/656852/) in 2020, found that even top performers counted **0.8%** of their disbursements as duplicate or erroneous, and bottom performers 2%. The ACFE numbers are correlations among organizations that were defrauded, and the APQC figures are older and count payments by number. Neither proves what an agent will save you. Both say the misses sit in places a full read reaches. ## The cost moves to the exception queue The model bill for a 500-person company will be tens of dollars a month. The people bill depends on how many items the checks flag and how long each takes to clear. If 2% of 11,000 monthly items reach a person and each takes ten minutes, that is 37 hours a month, close to a quarter of a full-time role. If the checks are badly tuned and flag 10%, the queue needs a full person and soon gets ignored. Every design decision should be judged by what it does to that number. The rule we use sorts every item into three buckets. Clear means the agent found nothing and the item proceeds. Review means a person looks before the item moves on, and the agent attaches the documents and its stated reason, so the reviewer reads one screen. Stop means the payment or reimbursement is held until a named owner decides, for a short list of findings such as a bank account that differs from the vendor master or an invoice that matches one already paid. The agent never moves an item out of review or stop by itself. The controls around an agent that acts in your ledger, its access rights, change management and sign-off, are the subject of [an earlier note](https://vandam.ai/notes/agents-and-internal-controls). The plan here keeps the agent in a narrower role. It reads and flags. People pay, approve and correct. ## Which populations to check first Start where the money leaves and the documents are already digital. **1. Accounts payable invoices.** For each invoice the agent checks whether it duplicates one already paid, including near-duplicates with a changed invoice number or amount; whether price and quantity match the PO and receipt; whether the remittance details match the vendor master; whether the vendor was created recently or by the same person approving the invoice; and whether a run of invoices from one vendor sits just below an approval limit. **2. Expense report lines.** The agent reads each receipt against the claim and the policy: amount and date match, merchant type allowed, the same receipt not claimed twice or by two employees, meals within limits for the city, and mileage consistent with the stated trip. Expense schemes ran longest in the ACFE data. **3. Vendor master changes and credit memos.** Every change to bank details, address or payment terms, and every credit memo issued to a customer, gets read against its supporting request. Volumes are small and the risk per item is high. **4. Contracts against invoices.** Once a month the agent reads each active vendor contract against that month’s invoices from the vendor, looking for prices above the agreed rate, escalators applied early, and renewals coming up with notice periods about to lapse. **5. Support tickets.** The agent reads every ticket for refunds or credits promised by an agent outside policy, complaints that describe a safety or legal issue, and urgent issues sitting in a normal queue. It sits outside finance, so it needs an owner in customer operations. ## Prove the checks work before you trust them A cheap check that misses things is worse than a sample, because it creates confidence. Before the agent replaces anything, run it for one full month alongside the current sampling. Then seed the population: plant 20 known problems, such as a duplicate invoice with one digit changed, a receipt reused across two reports, or a bank detail change with no request behind it, and count how many the agent catches. Track two numbers every month afterward: the share of seeded items caught, and the share of flags that a reviewer rejects as false. When the first is high and the second is low enough that reviewers still read each flag, the sample can shrink to the items the agent has never seen before. Keep the sampling your external auditor relies on until you have agreed the change with them. Under the new PCAOB rules they will want to see how flagged items were investigated, so keep the agent’s reason, the reviewer’s decision and the outcome on every flag. ## What it costs a 500-person company each month My working assumptions for a company of 500, which you should replace with your own counts: 3,000 AP invoices, 2,500 expense lines, 350 vendor master changes and credit memos, 300 contracts and 5,000 support tickets. That is about 11,150 checks a month. At $0.00075 each, the first pass costs about **$8**. Sending one item in ten to Sonnet 5.5 for a second opinion at about two cents each adds about $22. The model bill is roughly **$30 a month**, and long contracts or scanned receipts that triple the token count would still leave it under $100. The real costs are the build, a few weeks of work to connect the ERP, expense tool and help desk and to write and test the checks, and the reviewer hours, around 37 a month at a 2% flag rate. ## What to do this quarter For the CFO and controller, the sequence is short. Pull last year’s AP invoices and run the duplicate and vendor-master checks over all of them as a one-time look back; anything recovered pays for the project. Name one owner for the exception queue and set a service level for each bucket. Run the agent beside your current sampling for a month and seed it with known problems. Then take the results to internal audit and your external audit partner and agree in writing which samples the full check replaces. When those four are done, add expense lines, then contracts, then tickets, one population per month. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # What a firm should keep inside once agents make coordination cheap. · VANDAM > Coase's 1937 theory of the firm, rerun now that agents cut the cost of coordination: what moves out, what comes back in, and what stays expensive. URL: https://vandam.ai/notes/coase-and-the-firm Contact: hendrik@vibelab.be # What a firm should keep inside once agents make coordination cheap. In 1937 Ronald Coase explained that firms exist because using the market costs money. Agents now make buying, building and internal coordination cheaper all at once, and I list the five comparisons I would rerun at a 300-person company. Hendrik De Winne 25 October 2026 7 min read Halfway through an essay about selling to software, [Brian Flynn](https://x.com/Flynnjamm/status/2023465136204419096), a San Francisco builder whose profile points to Katch, a startup that pays people to capture what agents need verified, stops to concede something: “Not all transaction costs are falling equally. Integration, compliance, and security review are still expensive.” Then he moves on, because his essay is about the costs that are falling. I think the sentence he moved past is the one a company of a few hundred people should plan around. His essay opens with an economist most managers have never read, and the argument only works if you know what that economist actually said. ## A question from a lecture in 1932 Ronald Coase presented the argument in a lecture in Dundee in October 1932. He was **21**. It appeared as “The Nature of the Firm” in [Economica](https://doi.org/10.1111/j.1468-0335.1937.tb00002.x) in November 1937, and in 1991 he received the Nobel Prize in economics “for his discovery and clarification of the significance of transaction costs and property rights for the institutional structure and functioning of the economy.” The puzzle was simple. Economists said prices coordinate an economy. Yet inside every firm, work is coordinated by people giving instructions, and some firms were enormous. If the market did all the coordinating, firms would have no reason to exist. In his [Nobel lecture](https://www.nobelprize.org/prizes/economic-sciences/1991/coase/lecture/) he put his answer in plain terms: “there were costs of using the pricing mechanism. What the prices are has to be discovered. There are negotiations to be undertaken, contracts have to be drawn up, inspections have to be made, arrangements have to be made to settle disputes, and so on. These costs have come to be known as transaction costs.” A firm, that little planned society, could only continue to exist if it performed its co-ordination function at a lower cost than would be incurred if it were achieved by means of market transactions and also at a lower cost than this same function could be performed by another firm. (Ronald Coase, summarizing his 1937 article in his [Nobel lecture](https://www.nobelprize.org/prizes/economic-sciences/1991/coase/lecture/), 1991) Read as a manager, that sentence is a price comparison. Payroll sits in-house and cleaning is contracted because, at some point, one was cheaper to organize inside and the other cheaper to buy. Nobody reruns the comparison very often. Agents change both sides of it at once, which is why the boundary of a firm does not simply shrink. ## Agents made buying cheaper Flynn’s essay is about the market side. An agent “can discover a service, check its price, and call it in a single HTTP roundtrip. No proposal. No demo.” His conclusion follows Coase directly: “When search costs collapse, the default shifts from ‘build it in-house’ to ‘buy it on the open market.’” He gives his own worked example. An agent researching which data service to use spends $0.10 to $0.50 and 10 to 25 seconds reasoning it through, while a specialized catalog answers for $0.01 to $0.02 in under 200 milliseconds. In his words, “7-50x cheaper and 50-100x faster.” The models themselves are bought this way now. [Balaji Srinivasan](https://x.com/balajis/status/1931671546113994769), author of The Network State, described it last year: once a model is part of your workflow, “it’s easy to swap in a new one,” so “you essentially have budget for AI contractors in each vertical.” ## Building got cheaper too The other side of the comparison moved just as fast, and it pushes the opposite way. [Aaron Levie](https://x.com/levie/status/2004654686629163154), CEO of [Box](https://x.com/levie/status/2004654686629163154), wrote in December about a ten-person services firm that never had custom software, because building it would have taken several people. “Now, someone on the team builds a prototype in a few days, proves out the value proposition in a matter of days.” The cost of doing something yourself fell, so work that used to be bought, or skipped, comes inside. Sometimes there is nothing to buy. [Nikita Bier](https://x.com/nikitabier/status/2102432368158245252), former head of product at [X](https://x.com/nikitabier/status/2102432368158245252), wrote in September that when X looked for bot detection, “there was not a single company that brought together all the latest technologies so we had to do it all in-house.” And sometimes the thing you bought becomes cheap to replace. [Greg Isenberg](https://x.com/gregisenberg/status/2105397216374403561), CEO of [Late Checkout](https://x.com/gregisenberg/status/2105397216374403561), which builds and runs a portfolio of small companies, put it bluntly the same month: plenty of companies “stay with outdated software vendors only because switching is painful, and once rewriting is cheap, that lock in starts to break!” Flynn sees the same pull back inward from where he stands: “As models get cheaper and more capable, some services get absorbed back into the agent.” The services that survive, he writes, have “proprietary datasets, real-time feeds, hardware-dependent computation.” So the boundary moves in both directions at once. Generic work goes out to the market, specific work comes back in, and what decides it is whether anyone else has what the task needs. ## The cost that falls fastest is inside the firm Coase’s comparison has a third term people forget: the cost of organizing inside. A large share of a company’s headcount exists to coordinate, which means passing work between people and keeping them aligned. > in a world where AI handles the glue work...handoffs, specs, coordination, the need for internal labels shrinks fast. > > Greg Isenberg, CEO, Late Checkout, June 2025, [Source](https://x.com/gregisenberg/status/1931352236531040409) In the same post Isenberg wrote that “most legacy titles were invented to make headcount legible to org charts, not customers,” and that “the future of work is flatter. fewer layers. fewer middle steps. more output.” The small end of the market shows what that looks like. In 2025 [John Collison](https://x.com/collision/status/1942978418603360459), co-founder of [Stripe](https://x.com/collision/status/1942978418603360459), interviewed the indie founder Pieter Levels about “how he’s making $3m a year as a one-person company.” This June [Nicolas Finet](https://x.com/nifinet/status/2067638000037790178), CEO of [Sortlist](https://x.com/nifinet/status/2067638000037790178), published how he runs prospecting, call prep, follow-up, no-show recovery and weekly reporting as five agents with one shared memory. His instruction is the Coase question in one line: “stop thinking in headcount and start thinking in jobs.” He reports the whole setup “runs for about $400 a month in tokens,” and is careful about the limit: the agents “do not close.” Larger firms are testing the same thing. Coinbase’s “one person teams,” which I covered in [the note on the new org chart](https://vandam.ai/notes/org-chart-redrawn), are the inside version of this: one person carrying work that used to need a handoff between three roles. ## Checking the work and settling disputes stay expensive This is where Flynn’s caveat matters. Go back to Coase’s list from 1932 and check each item against what agents do. Discovering prices is close to free for an agent. Negotiating, in Flynn’s model, moves into the protocol, with the price returned in the response. Contracts become allowlists and spending limits. The last two items on the list are the ones agents have barely touched. Inspection, meaning checking that the work was done right, is still hard, and Flynn’s own data shows it. - **2/44** Services with fully working endpoints in one sweep of an agent service catalog, cited by [Brian Flynn](https://x.com/Flynnjamm/status/2023465136204419096). - **53%** Of direct service calls that succeeded in the same data. - **87%** Success rate of the recommendation layer that chose between services. Settling disputes is the other one. When an agent buys the wrong thing, someone in the company still answers for it. Flynn expects enterprise buying to “re-optimize around ‘allowed, trusted, and auditable’ alongside ‘fast and cheap.’” Each of those three words is work that a person inside the company has to do, which means integration into the systems you already run, a security review before anything touches customer data, and someone who checks the output. There is one more cost Coase did not list, because in 1932 it lived in people’s heads: what the company knows about its own customers, exceptions and past decisions. A vendor cannot sell you that, and an agent without it makes generic decisions. I wrote about why that record compounds in [the note on company context](https://vandam.ai/notes/context-compounds). For the boundary question it means the specific knowledge stays inside, and it needs an owner. ## What I would rerun at a 300-person company Coase’s test is still the right one. The prices in it changed, so the answers from the last time anyone ran it are out of date. These are the five comparisons I would put in front of a leadership team. 1. **Every outsourced task with a clear output.** Price it three ways: the current supplier, an AI-native supplier that sells the finished work, and one of your own people running agents. The third option did not exist when the contract was signed. 2. **Every tool you bent your process around.** If the team uses a narrow slice of a large product, ask what the narrow version would cost to build and maintain. Keep the systems of record, and question the rest. 3. **Every role whose main job is routing.** Handoffs, status reports and specs for other teams are the internal coordination costs falling fastest. Redesign those roles with the people in them, before anyone draws a new chart. 4. **Everything only your company has.** Customer history, exception rules, pricing logic and supplier relationships are what agents need and no vendor can supply. That work moves inside, and someone should own it by name. 5. **The costs that do not fall.** Budget for integration, security review and checking agent output as permanent line items, with an owner, before buying more agents. Most of the money in this exercise sits in the first and third items. Most of the risk sits in the fifth, because a company that buys cheaply and skips the inspection has only moved its transaction costs to the day something goes wrong. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # What your company knows becomes the asset once agents do the work. · VANDAM > Once agents do the work, a company's decisions and the reasons behind them become its asset. Why most firms lose them, and four things to capture first. URL: https://vandam.ai/notes/context-compounds Contact: hendrik@vibelab.be # What your company knows becomes the asset once agents do the work. Agents only know what someone wrote down, and most companies record the outcome of a decision while losing the reason, the exception and the person who said yes. This note explains why that record compounds and lists the four things I would start capturing this month. Hendrik De Winne 24 October 2026 7 min read A renewal agent proposes a **20%** discount. Company policy caps renewal discounts at 10% unless someone approves a service-impact exception. The agent pulls three severe incidents from the on-call system, an open “cancel unless fixed” escalation from the support desk and last quarter’s renewal thread, where a VP signed off on a similar exception. It sends the request to Finance, and Finance approves. When the work is done, the CRM holds one fact: “20% discount.” The example comes from an essay by [Jaya Gupta](https://x.com/JayaGup10/status/2003525933534179480), a partner at [Foundation Capital](https://foundationcapital.com/team/), written with Ashu Garg, a general partner at the same firm, and published in December 2025. I keep coming back to it because I see the same thing in every company I work with. The systems keep the outcome. The incidents, the threat to cancel, the precedent and the person who said yes all stay in a chat thread, a call and somebody’s memory. Next year, when the same customer asks for the same discount, nobody can tell whether 20% was a one-off or the new rule. For most of the history of the modern company, that loss was tolerable. The person who granted the exception was still around, and the colleague handling the next one could walk over and ask. Once agents do a large share of the work, that hallway stops existing. An agent has what it can read, and nothing more. ## The record keeps the answer and loses the reason Gupta and Garg call the missing material decision traces: “the exceptions, overrides, precedents, and cross-system context that currently live in Slack threads, deal desk conversations, escalation calls, and people’s heads.” They separate two kinds of knowledge an agent needs. Rules “tell an agent what should happen in general.” Decision traces “capture what happened in this specific case,” including which policy applied, who granted the exception and which earlier case it followed. Their list of what goes missing reads like a tour of any mid-sized company. A sales rule that exists only as habit, such as an extra 10% for healthcare customers “because their procurement cycles are brutal.” A deal structured to match one from last quarter, with nothing linking the two. A support lead who checks revenue, open escalations and a worried chat thread before escalating a ticket, after which the ticket says only “escalated to Tier 3.” A VP who approves a discount on a video call, so the record shows the price and never the approval. The reasoning connecting data to action was never treated as data in the first place. (Jaya Gupta and Ashu Garg, [Foundation Capital](https://x.com/JayaGup10/status/2003525933534179480), December 2025) Their essay is written for founders, and its conclusion is about which software companies will own this layer. My interest is narrower. Whoever sells the software, the reasons behind your decisions belong to your company, and today most companies throw them away by default. ## Agents only know what someone wrote down [Eli Mernit](https://x.com/mernit/status/2021324284875153544), founder of [Beam](https://www.beam.cloud/), wrote the most literal version of this idea in February, in a post called “Your Company is a Filesystem.” He imagines a law firm as a set of folders: a new case is written to /cases, logged time goes to /billing/time-sheet, and “permissions naturally map to seniority in the org chart.” An agent working in that firm can read and write whatever it needs. The trouble he names is that real companies keep things everywhere: “Invoices are in Quickbooks, emails are in Outlook, proposals live in Sharepoint.” He also concedes the harder point, that “many work streams are codified in people’s heads - not in JSON files.” Connecting the systems is the part most AI projects already budget for. The heads are where projects stall, and the most detailed evidence I know of comes from a company that has gone further down this road than anyone. In its October study of its own workforce, [Anthropic](https://www.anthropic.com/research/employment-at-anthropic) reports that written standing instructions for its agents grew from a few hundred in July 2025 to **8,500** a year later. Its conclusion about everything that was not written down is blunt: “Private work, in unlogged meetings and direct messages, is invisible, confusing and discouraged in an AI-heavy workplace.” And the reason given is the same one Gupta gives: “Keeping human work visible across the company ensures autonomous agents reason from fresh context.” Eight and a half thousand instructions in one year means people spent real working time writing down how they wanted things done. That writing is now part of how the company runs. It is also an asset the company did not have two years ago, and nobody outside the company has a copy of it. ## Context is what stays when people and models change [Greg Isenberg](https://x.com/gregisenberg/status/2086459363842048422), CEO of the studio [Late Checkout](https://latecheckout.studio/), sketched the new org chart in August as a small layer of people on top “doing strategy, creative, and judgment,” with a wide layer of agents underneath. His line about what that means for managers is the one I quote most: “Managing people became managing agents, and managing agents is really just managing context.” > Both humans/agents come and go, the context is the thing that compounds and can’t be copied. > > Greg Isenberg, CEO, Late Checkout, August 2026, [Source](https://x.com/gregisenberg/status/2086459363842048422) The compounding is concrete. Every model a company uses will be replaced within a year or two, and it should be. The agents get rebuilt on whatever comes next. People change roles and leave. What carries over is the record of how the business decides: which customers get exceptions, which suppliers are trusted and why, where the line sits between a call a junior can make and one that goes to the owner. Isenberg made the same point about agent software [a month later](https://x.com/gregisenberg/status/2099202686377742576): “every correction a human makes becomes a rule the harness keeps.” A company that captures those corrections gets better each quarter. A company that lets them evaporate starts over each time a person leaves or a system is replaced. Compounding works in both directions, and I would be careless to skip that. A record full of workarounds compounds the workarounds. Ahmad Kakar, CEO of Euclid, told [Annelies Gamble](https://x.com/AnneliesGamble/status/2069814714214023186) of Zetta Venture Partners that a considerable part of tribal knowledge is “workarounds people built for non-existent or broken systems,” and that the job is “deciding what was real judgment versus scar tissue.” Stale context does similar damage. An instruction nobody owns stays in force long after the reason for it is gone, and an agent follows it with complete confidence. So the record worth having is curated, owned and dated, which means someone decides what stays out of it. ## What I would start writing down this month Gupta and Garg write that “none of this requires full autonomy on day one,” and the same goes for software. A company can start with a habit and a shared place for it to land. These are the four things I ask teams to capture first, in this order, because each one is cheap to record at the moment it happens and expensive to reconstruct later. 1. **Decisions with their reasons.** One paragraph per decision that changes how work is done: what was decided, what was considered and rejected, and what fact would make you decide differently. 2. **Exceptions, with who approved them and why.** The discount above the cap, the late payment that was waived, the supplier allowed in without the usual checks. Exceptions are where the real policy lives, and they are the first thing an agent gets wrong. 3. **Owners for every rule and instruction.** A name and a review date on each one, so that a rule which no longer holds has someone whose job it is to remove it. 4. **Answers that cite their source.** When someone, or something, answers a question about how the company works, the answer links to the document or decision it came from. That is how people learn to trust the record, and how they spot the parts that are out of date. This is also how we build at VANDAM. A [Company Brain](https://vandam.ai/company-brain) answers everyday questions from the documents, emails and decisions a company chooses, with each answer linked to its source, so the third and fourth items have somewhere to live. A [Judgment Vault](https://vandam.ai/judgment-vault) starts with one person whose decisions keep landing on their desk, works through their past cases until the reasoning is on paper, and separates real judgment from old workarounds before anything reaches an agent. Both start from the same view: the reasons are the valuable part, and they have to be collected while the people who hold them are still in the room. If you want a quick test of where your company stands, take the last three exceptions someone granted this month and try to find out why each one was approved without asking the person who approved it. In most companies I have seen, you will find the outcome in a system within a minute and the reason nowhere at all. That gap is the first thing to close before agents start making those calls on their own. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # The AI line in your budget should be priced per finished task. · VANDAM > AI return on investment as a CFO should see it: why budgets move from licenses to tokens and outcomes, and how to measure cost per completed task. URL: https://vandam.ai/notes/cost-of-the-i Contact: hendrik@vibelab.be # The AI line in your budget should be priced per finished task. AI spend is moving from seats and headcount to tokens and outcomes, while most finance teams still count licenses. Intercom, Box, Podium and Varick show what the cost side of AI looks like, and I name the three numbers I would ask a CFO to own. Hendrik De Winne 25 October 2026 7 min read The pricing page of [Intercom](https://www.intercom.com/pricing) shows two ways of paying for software on one screen. The help desk on the Advanced plan costs **$85** per seat a month, billed annually. Right below it, the company’s AI agent, Fin, costs **$0.99** per outcome. An outcome is counted when a customer confirms the issue is resolved, stops asking for help, or Fin completes a workflow. The first price buys a person access to a tool. The second buys a piece of work that got done. Most finance teams I talk to still budget AI the first way, as licenses on the software line, while the vendors have started to bill the second way. That mismatch is where most of the confusion about AI return on investment comes from. In April, [Alex Lieberman](https://x.com/businessbarista/status/2047488697885413391), founder of the AI transformation firm Tenex and co-founder of Morning Brew, co-led an AI training for 20 executives from what he called one of the largest financial institutions in the world. The first “aha moment” on his list was about money: “Historically, leaders had to think about allocating budget for headcount and software. Now, token budget must be a serious consideration.” His questions for that budget are the ones I would hand any CFO: “Do you set limits per employee? How do you measure the ROI on your token budget?” ## Box’s CEO says to watch the I in ROI The clearest argument for why this budget line will grow came from [Aaron Levie](https://x.com/levie/status/2004654686629163154), CEO of Box, in an essay he called “Jevons Paradox for Knowledge Work.” William Stanley Jevons noticed in the nineteenth century that more efficient use of coal led to more coal being burned, because cheaper coal found new uses. Levie applies that to the work agents now do, from reviewing contracts to running market research. > The mistake that people make when thinking about ROI is making the “R” the core variable, when the real point of leverage is bringing down the cost of “I”. > > Aaron Levie, CEO, Box, December 2025, [Source](https://x.com/levie/status/2004654686629163154) His point is that a small company has always had to choose between a better website, a support desk and a cleaner finance process, because each one cost people it did not have. Once the cost of doing a task drops far enough, those trade-offs loosen. In his words, “every business in the world has access to the talent and resources of a Fortune 500 company 10 years ago.” He expects demand for many kinds of work to go up “10X or 100X,” and he expects most future tokens to be spent on “things we don’t even do today as workers.” For a finance team, that has an awkward consequence. If Levie is right, total AI spend should rise, and a rising bill tells you very little about whether the money is working. The useful number is the unit cost: what one finished piece of work costs now, compared with what it cost before. ## Podium sells against the staffing budget [Eric Rea](https://x.com/ericwilliamrea/status/2013980401635582277), co-founder and CEO of [Podium](https://www.podium.com), described in January how his company went from selling software to local businesses to selling them a worker that runs it. Podium already helped more than 60,000 local businesses handle leads and customer messages. Then, he wrote, “one reality became impossible to ignore: Our customers’ biggest constraint isn’t software. It’s staffing.” - **75%** Annual staff turnover among Podium’s customers, as stated by [Podium](https://x.com/ericwilliamrea/status/2013980401635582277). - **30%** Share of their leads that arrive after hours, per the same post. - **10,000+** AI agents Podium says it has live in production. These are the company’s own numbers, from a CEO announcing “$100M+ in AI agent ARR” in under two years, so I treat them as a claim. The logic behind them still matters for anyone holding a budget. Podium’s agent, which it calls Jerry, qualifies and schedules leads around the clock inside Podium’s own software. A business owner judges that against what a person on the front desk costs. The money comes out of the staffing line, and the owner measures it the way he would measure a hire: are the leads answered, and do they turn into appointments. ## Counting licenses hides where the money goes Inside large companies the pattern runs the other way. [Vas](https://x.com/vasuman/status/2085806422072418632), CEO of [Varick Agents](https://varickagents.com), wrote in August about an executive who had just signed “an eight-figure commitment for the year” in enterprise licenses. The result: “Roughly 10% of people burn 90% of the tokens.” Share of people 10% Share of tokens 90% One company’s enterprise AI rollout, as described by [Vas, Varick Agents](https://x.com/vasuman/status/2085806422072418632) , August 2026. He then ran the sum most budgets skip. If everyone else used AI the way the top tenth does, “spend goes up roughly 10x. $10M of commitment becomes $100M, and your best-case scenario is now your worst.” The license count says the rollout succeeded. It says nothing about which of those tokens produced anything. Vas names the measure he would want instead: “what is their ROI per token used?” Optimized model routing is the new salary banding. You wouldn’t pay executive salary for data entry (Garrett Lord, CEO of [Handshake](https://x.com/GarrettLord/status/2068754262440767500), June 2026) Lord’s comparison is the one that made the problem click for me, because finance already knows how to do salary bands. You match what you pay to the difficulty of the work, and you check the output. He adds the condition that keeps it honest: routing to a cheaper model “only works if you have the evals to know whether a cheaper model can actually deliver,” and companies that over-optimize for cost “pay for it in quality.” [Gartner](https://cpapracticeadvisor.com/2026/05/28/gartner-says-cfos-must-stop-mistaking-finance-ai-deployment-for-value-creation/184094) put the same warning to CFOs in May: they “must not mistake activity for impact.” I covered that survey in the note on [why our AI team hires from HR and finance](https://vandam.ai/notes/ai-needs-hr-and-finance). ## Three numbers I would ask finance to own If the budget is moving from seats to tokens and outcomes, the measures should move with it. These are the three I would put in a monthly report, for each workflow where AI does part of the work. 1. **Cost per completed task.** Take everything the task now costs: tokens, the licenses it uses, the hours people spend reviewing and fixing the agent’s output, and the rework that comes back later. Divide by the number of tasks finished to an agreed standard, and compare it with the same task before. Intercom’s definition of an outcome is a good template, because it says exactly when a piece of work counts as done and charges nothing when it is not. 2. **Share of the work that is manual, hybrid or automated.** Vas suggests boards should see this split, and I argued in [the note on adoption and diffusion](https://vandam.ai/notes/adoption-vs-diffusion) that it beats any login count. It shows whether the work itself is changing, one workflow at a time. 3. **Quality against the human baseline.** Lieberman’s working model for AI ROI makes this a hard condition. In his resume-screening experiment, AI-picked and human-picked candidates go blind to the hiring manager, and the experiment only counts as a success if time saved is above zero and the “AI Average Quality Score ≥ Human Average Quality Score.” A falling cost per task with falling quality is a loss that shows up later. With those three in hand, the decision rule gets simple. Vas, quoted by [Annelies Gamble](https://x.com/AnneliesGamble/status/2069814714214023186) of Zetta Venture Partners, put it as a gate: “A workflow with a 10x ROI should get prioritized. A workflow with a 2x ROI maybe, maybe not. And a workflow with a sub 1x ROI should be left alone.” You can only apply that gate if you know what the task costs. ## Small spend can stay loose The strongest objection is that this is too much accounting for something this cheap and this new. Lieberman himself argues that below a certain level of spend, ROI “can be vibes-based through conversations with users,” because the goal at that stage is to “remove friction & empower people to play with the technology.” He is right about that. A finance team that demands a cost-per-task study before anyone may try an agent will get a tidy report and no learning. Jevons also cuts against tight caps. If the cheap tokens are the ones that start the contract review nobody used to do, a hard spend limit per employee can kill the work with the best return before it shows up. So I would split it the way Lieberman does. Let small spend stay loose, and keep the three numbers for the workflows where the money is real: the ones on the eight-figure contract, the ones that replace a hire you would otherwise make, and the ones a board will ask about. For those, the license count can stay in the appendix. If you run finance, pick the single workflow with the largest AI bill this quarter and write down what one finished task cost last month, what it cost a year ago, and who checked the quality both times. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Delete the step first and give the agent whatever survives. · VANDAM > Musk’s five steps put automation last. Why AI makes automating first cheap and invisible, and how to run accounts payable in the right order. URL: https://vandam.ai/notes/delete-the-step-first Contact: hendrik@vibelab.be # Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. Hendrik De Winne 26 October 2026 9 min read On April 13, 2018, with Model 3 production far behind plan, [Elon Musk](https://x.com/elonmusk/status/984882630947753984) answered a Wall Street Journal reporter on Twitter: “Yes, excessive automation at Tesla was a mistake. To be precise, my mistake. Humans are underrated.” The same day, [TechCrunch](https://techcrunch.com/2018/04/13/elon-musk-says-humans-are-underrated-calls-teslas-excessive-automation-a-mistake/) carried what he had told CBS about the Model 3 line: “we had this crazy, complex network of conveyor belts….And it was not working, so we got rid of that whole thing.” Three years later, walking the SpaceX site in Texas with the YouTuber Tim Dodd, Musk gave the best example from that period. Fiberglass mats sat between the Model 3 battery pack and the floor pan, and the robot that placed them was choking the battery line. He first tried to make the robot faster, then simplified the gluing, and only then asked what the mats were for. The battery safety team said noise and vibration. The noise and vibration team said fire safety. Tesla then built one car with the mats and one without, put a microphone in each, and nobody could hear a difference. The mats were deleted, and with them, in Musk’s words, “this two million dollar robot cell,” according to the [Starbase tour video](https://www.youtube.com/watch?v=t705r8ICkRw&t=1325s). That robot was at least visible. It sat on the line, it slowed production, and the CEO ended up living next to it until someone asked the right question. The AI version of the same mistake is quieter. An agent that resolves invoice exceptions costs a few cents a document and never jams anything. It clears the mismatches so fast that nobody sees them pile up, and the pile was the only sign that a step upstream should not exist. Most companies buying agents today start at the last step of Musk’s method, and AI is the first technology that makes starting there cheap enough that nobody is ever forced to go back. - **$2M** Robot cell Tesla bypassed once a test showed the mats it placed did nothing, per Musk on the [Starbase tour](https://www.youtube.com/watch?v=t705r8ICkRw&t=1480s), 2021. - **18.4%** Average invoice exception rate in accounts payable, per [Ardent Partners](https://payablesplace.ardentpartners.com/2026/01/state-of-epayables-part-nine-ap-benchmarks-and-best-in-class-performance/), January 2026. - **80%** Share of Claude Code’s system prompt [Anthropic](https://x.com/trq212/status/2080710971228918066) removed for its newest models, with no measurable loss on its coding evaluations. ## Musk’s five steps only work in order On the same tour, Musk laid out the process he wants SpaceX to “implement rigorously.” Step one is to make your requirements less dumb. Step two is to try very hard to delete the part or process. Step three is to simplify or optimize. Step four is to accelerate cycle time. Step five is to automate. He was blunt about which step people reach for first: “possibly the most common error of a smart engineer is to optimize the thing that should not exist.” And he admitted the Model 3 line was his own case of running the list backwards: he automated, accelerated, simplified, and deleted last. Two rules inside the first two steps matter more to a COO than the list itself. The first is that every requirement “must come with a name, not a department,” because a department cannot be asked why a rule exists and a person can. Otherwise, he said, you obey a rule an intern made up two years ago. The second is a test for whether you deleted enough. If you are not adding something back in about 10% of the time, you are not deleting enough. His reason: “the bias tends to be very strongly towards let’s add this part or process step in case we need it.” Office work has the same bias, with less friction against it. A second approval, an extra field on a form or a weekly report gets added after one bad quarter and never removed, because nobody’s job is to remove it. ## An agent makes the backward order cheap and quiet Two things make the backward order more tempting with AI than it ever was with robots. The first is cost and visibility. Tesla’s robot cells cost millions and slowed a line the whole company was watching, so someone had to ask what they were for. An agent costs a subscription and runs inside software nobody looks at. Step five is also the only step a vendor can sell you. Nobody invoices you for deleting an approval. The second matters more. Exceptions are how a broken process shows itself. In accounts payable, an exception is an invoice that does not match its purchase order or receipt, and the average AP department sees one on **18.4%** of its invoices, according to [Ardent Partners](https://payablesplace.ardentpartners.com/2026/01/state-of-epayables-part-nine-ap-benchmarks-and-best-in-class-performance/). Ardent calls exceptions the biggest single reason AP benchmarks are not better, and puts the average cost to process one invoice at **$9.84**. Every exception has a cause upstream: a buyer who issued a PO with the wrong unit of measure, a vendor master record with two addresses, a price nobody updated after the contract changed. When a clerk handles those exceptions, the queue grows, the clerk complains, and somebody eventually goes upstream to fix the PO. When an agent handles them, the queue stays empty. The cause stays in place, the agent keeps getting paid to clear its effects, and the dashboard reports a fast, cheap process. Michael Hammer made the same point about Ford’s payables department in [HBR](https://hbr.org/1990/07/reengineering-work-dont-automate-obliterate) in 1990, a case I walk through in [an earlier note](https://vandam.ai/notes/ten-steps-one-is-ai): helping the clerk investigate mismatches faster was the weaker choice, and the better one “was to prevent the mismatches in the first place.” An agent is the most efficient investigator of mismatches ever built. ## Accounts payable, run in Musk’s order Here is the method on one process. The company is an illustration, and no client is behind it: a 700-person maker of commercial HVAC equipment, with a few thousand vendor invoices a month. Invoices arrive as email PDFs. A clerk keys them into the ERP and runs a three-way match against the purchase order and the receiving record. Mismatches go to an exceptions queue. Invoices over $5,000 need two approvers. Payments run once a week. The proposal on the CFO’s desk is an agent that reads invoices and resolves exceptions. **1. Question every requirement, and put a name on it.** List every rule the process follows and write next to each one the person who owns it today. Who set the $5,000 threshold for a second approver, and why does an invoice that already matches an approved PO need one? Why does the company wait for an invoice at all from vendors whose PO and receipt already agree? Why do 40 general-ledger codes get assigned by hand when the monthly report uses 12? The rule we use: a requirement with no current owner who will defend it goes on the delete list. Some requirements will have strong owners. Segregation of duties stays. The call-back to confirm a vendor’s new bank details stays too, because business email compromise cost US victims **$3.0 billion** in 2025, according to the [FBI’s IC3](https://www.ic3.gov/AnnualReport/Reports/2025_IC3Report.pdf), and that is how most of it gets in. **2. Delete.** For the top vendors by volume, pay on receipt against the PO price and stop asking for invoices, which procurement calls evaluated receipt settlement. Drop the second approval on PO-backed invoices. Stop hand-coding fields no report uses. Then count the exceptions that disappeared, and apply Musk’s test: if, after a quarter, you never had to restore a single step, you were too timid. **3. Simplify.** Match on the few fields that decide whether to pay: part number, quantity, price. Keep one vendor master, one intake channel for invoices and one approval path. **4. Accelerate.** Move the payment run from weekly to daily. Put an age limit on every open exception and route it back to the buyer who created the PO, so the fix happens upstream. **5. Automate.** Now the agent gets what survived: true exceptions, invoices without a PO, vendor questions. It is a smaller job, and easier to audit. And because the noise is gone, each exception it still sees points at a cause somebody can fix. Keep the exception rate on the COO’s dashboard after go-live, broken out by cause, so the agent’s speed never hides it. Order-to-cash takes the same sequence. Ask who set the credit hold limit and why it is a fixed dollar amount. Delete the retyping between quote and sales order. Simplify the pricing exceptions sales can grant. Speed up cash application. Only then put an agent on the disputes that are left. ## Deleting a step needs someone allowed to delete it The hard part of steps one and two is authority. The second approval belongs to the controller, the GL codes to FP&A, the PO template to procurement and the credit limit to sales. Each owner can defend their rule, and nobody in the chain can delete another department’s step. Box CEO [Aaron Levie](https://x.com/levie/status/2096279756925026733) quoted the clearest description of the result I have seen: you “‘apply AI’ on what you currently have, and you end up with faster sh*t.” I wrote about why that mandate belongs with HR and finance in [another note](https://vandam.ai/notes/ai-needs-hr-and-finance). So before choosing a vendor, the CEO has to give one named person the right to delete steps across every department the process touches, with a deadline and a written list of what is off limits: statutory controls, audit trail, tax. Musk’s rule works in both directions here. Each requirement needs a name, and so does the deletion. ## Run the five steps again every time the model changes Automating is also where new requirements pile up. Every agent comes with instructions, guardrails, approval prompts and a library of rules for edge cases, written to steer a model that made mistakes. Those are requirements too, and they age faster than any approval threshold. In July, Thariq, who works on Claude Code at Anthropic, [wrote](https://x.com/trq212/status/2080710971228918066): “We removed over 80% of Claude Code’s system prompt for models like Claude Opus 5 and Claude Fable 5 with no measurable loss on our coding evaluations.” His explanation: “we found that we were over-constraining Claude Code.” Rules that were once needed to avoid worst cases could be deleted, and the model used context and judgment instead. That is a lab tuning its own product, but the timing carries over. Every major model release is a step-one event for the instructions you wrote around the last one. Give each standing instruction an owner and review the set at every model change, the way you would review controls after an ERP upgrade. This method has limits. It comes from manufacturing, where a mat either changes the noise in the cabin or does not. Office processes carry legal and control requirements that a test cannot settle, and those stay. And Tesla’s factories are heavily automated today, which fits Musk’s own reading: his lesson was about the order of the steps, with automation placed last. ## What to do this quarter Pick the one process where an AI proposal is already on the table, most likely AP, order-to-cash or customer service. Before signing, ask the vendor to price the agent on the process as it will look after deletion, and give yourself six weeks to get it there. In week one, name one person with authority across every department the process touches. In weeks two and three, list every requirement with its owner and the year it was introduced, and mark the ones nobody will defend. In weeks four and five, delete them, simplify what is left, and measure the exception rate and cost per transaction before and after. In week six, write the agent’s brief against the smaller process, with the exception rate as a metric it reports. Then compare the two quotes. The difference is what the company would have paid every month to run steps it should have deleted. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. ## Fix a mistake once by turning every correction into a rule the agents read. Anthropic's data team took Claude from 21% to above 95% accuracy with written skills, and now turns every correction a colleague makes into a rule and a test for the person who owns that domain. I set out that loop for a mid-sized company, with where the rules live, who owns them, and the one number a COO should track: how often the same mistake comes back. --- # Europe can lose the model race and still win on AI. · VANDAM > Europe's case for AI diffusion over frontier models: Ding, Draghi, Apply AI, Eurostat, the ECB, and the access objection, read at company level. URL: https://vandam.ai/notes/europe-diffusion-bet Contact: hendrik@vibelab.be # Europe can lose the model race and still win on AI. At a keynote this week I heard the case that Europe should stop chasing frontier models and put everything into getting AI used in every company. I checked the evidence behind it, read the strongest objection, and worked out what it means for a single firm. Hendrik De Winne 24 October 2026 7 min read At a keynote I attended this week, the speaker made a case I have heard in pieces for a year but rarely as one argument. Europe will not build the best AI model. It should stop spending political energy on trying. The race it can win is the slower one: getting every company, including the small and boring ones, to use AI and to reorganize how it works around it. It is an attractive argument for someone in my line of work, which is a good reason to distrust it. So I went back to the sources it rests on, checked the numbers, and read the best case against it. Most of it holds. One objection is serious, and it changes how I would set up AI inside a company. ## Inventing matters less than spreading The intellectual root is a book. [Jeffrey Ding](https://jeffreyjding.github.io/), an assistant professor of political science at George Washington University, argues in [Technology and the Rise of Great Powers](https://press.princeton.edu/node/51524) (2024) that countries pulled ahead in past industrial revolutions through “institutional adaptations oriented around diffusing technological advances throughout the entire economy.” The country with the first breakthrough often lost to the country that put the breakthrough to work in more places. The same year, Mario Draghi, former president of the European Central Bank, delivered his [report on European competitiveness](https://commission.europa.eu/topics/eu-competitiveness/draghi-report_en) to the Commission. His diagnosis of why Europe fell behind the United States after the mid-1990s reads like a case study for Ding. Europe failed to capitalize on the internet “both in terms of generating new tech companies and diffusing digital tech into the economy.” The gap sits largely in tech: “if we exclude the tech sector, EU productivity growth over the past twenty years would be broadly at par with the US.” And within those industries, the report points to “slow technology diffusion” as the reason the gap was largest exactly where productivity grew fastest. Draghi wanted both things, to be fair to him. He wrote that Europe needs to lead in new technologies and also “integrate AI into our existing industries so that they can stay at the front.” He also saw AI as “a window” for Europe “to redress its failings in innovation and productivity.” The keynote version of the argument keeps the second half and lets go of the first. ## Brussels turned it into a strategy On October 8, 2025, the Commission adopted the [Apply AI Strategy](https://digital-strategy.ec.europa.eu/en/policies/apply-ai). The name says what it is about. It targets ten industry sectors plus the public sector, from healthcare and manufacturing to energy and agri-food, and asks for “an ‘AI first policy’ where AI is considered as a potential solution whenever organisations make strategic or policy decisions.” The European Digital Innovation Hubs, the network meant to help smaller firms, were turned into Experience Centres for AI. The numbers show how far there is to go. - **20%** EU enterprises with ten or more staff using at least one AI technology in 2025, up 6.5 points in a year, per [Eurostat](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises). - **55:17** Share of large firms against small firms using AI in the EU, 2025, per [Eurostat](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises). - **7%** Euro area firms that use AI significantly, per the [ECB](https://www.ecb.europa.eu/press/key/date/2026/html/ecb.sp260323_1~1e06784a89.en.html), March 2026. Eurostat counts **19.95%** of EU enterprises using AI in 2025, against roughly 13.5% the year before. That is fast growth from a low base. The gap by size is the part I keep coming back to: **55%** of large enterprises, **17%** of small ones. And among firms that considered AI and still did not use it, the most common reason was a lack of relevant expertise, at 70.89%. Unclear rules and privacy came after that. The ECB number is the sobering one. About two-thirds of euro area firms in its survey say their employees use AI. Only **7%** use it significantly. Most of what gets counted as adoption is light, occasional use. ## The central bank put a number on the bet In March, [Philip Lane](https://www.ecb.europa.eu/press/key/date/2026/html/ecb.sp260323_1~1e06784a89.en.html), member of the ECB’s Executive Board, walked through ECB staff scenarios for what AI does to productivity. The range was wide, and he said it is “primarily driven by the speed of diffusion.” Fast adoption could add around **0.3 to 0.4** percentage points a year to total factor productivity over the coming decade. Slow adoption, the pace of earlier general-purpose technologies, gives about **0.2**. He concluded that policies focused on diffusion, training and SME adoption “could be especially beneficial.” If AI’s productivity dividend is adoption-dependent, then the diffusion trajectory in Europe becomes the decisive variable. ([Philip Lane](https://www.ecb.europa.eu/press/key/date/2026/html/ecb.sp260323_1~1e06784a89.en.html), ECB Executive Board, March 2026) The difference between those scenarios is one or two tenths of a point a year, which sounds small. Compounded over a decade across the whole euro area, it decides whether AI shows up in the productivity statistics as a rounding error or as the first real movement in years. The variable that decides it is how fast ordinary firms change, which is the keynote’s argument in a central banker’s words. ## The sharpest version came from Berlin In June, the Berlin think tank [Dezernat Zukunft](https://dezernatzukunft.org/en/publications/europes-ai-strategy-the-wrong-reflex) published the bluntest form of the argument, written by its founder and director Philippa Sigl-Glöckner with economist Paul Görlich and data scientist Alexander Marx. They recommend Ding’s book at the bottom of the piece, which tells you where the thinking comes from. > The fight over US foundation models, and over the infrastructure to develop them, is the wrong one. If Europe wants to grow economically and geopolitically stronger in AI, it should bet everything on rapid diffusion and accessibility. > > Philippa Sigl-Glöckner, Founder and director, Dezernat Zukunft, June 2026, [Source](https://dezernatzukunft.org/en/publications/europes-ai-strategy-the-wrong-reflex) Their reasoning has two legs. The first is that “the business model of America’s AI labs depends on the European market,” so the fear of being cut off is overdone. The second is that “models are most likely becoming a commodity,” with open models close behind the closed ones at a fraction of the price. From there they go after the EU’s plan to build five AI gigafactories for €20bn, which they say could become “the most expensive white elephant of European industrial policy.” Their alternative is software that makes AI usable, enough inference capacity in Europe, governments as anchor customers, and bargaining power where Europe already has it, in chipmaking equipment. ## The strongest objection is about access Eleven days later, [Foreign Policy](https://foreignpolicy.com/2026/06/29/europe-ai-superpower-eurostack-anthopic-mistral-openai/) ran a piece by [Thorsten Benner](https://gppi.net/team/thorsten-benner), co-founder and director of the Global Public Policy Institute in Berlin, and Jakob Hensing, its head of political economy. They agree Europe will not become an AI superpower. Europe has 5 percent of the world’s computing power and its best model trails the American and Chinese leaders. Where they part ways with the diffusion camp is on how safe it is to rely on someone else’s frontier. Their trigger was concrete. In June the U.S. government imposed export controls on Anthropic’s most advanced models, Mythos 5 and Fable 5. “Even if the U.S. government were to lift these restrictions soon,” they write, “it is obvious that this can happen again at any time.” > It is absolutely essential that Europe secures access to the leading U.S. models. Second-best models would likely fall short of effectively guarding against critical cyber- and other security risks. > > Thorsten Benner and Jakob Hensing, Director and head of political economy, GPPI, in Foreign Policy, June 2026, [Source](https://foreignpolicy.com/2026/06/29/europe-ai-superpower-eurostack-anthopic-mistral-openai/) I take this seriously, because it is the one place where the diffusion argument has a single point of failure. A strategy built on using other people’s models only works while you can get them. Benner and Hensing’s answer is close ties with U.S. industry, a strong position in industrial AI and robotics, and coalitions with other middle powers. Dezernat Zukunft’s answer is that American labs need European customers too much to shut them out for long. Both may be partly right. Neither side argues that Europe should try to out-train Anthropic or OpenAI. The live disagreement is about how much insurance the diffusion bet needs. ## What diffusion means inside one company Everything above is about continents. The bet only pays out inside individual firms, and that is where the Eurostat numbers become a to-do list. Most small and mid-sized companies are not using AI at all. Of those that looked and stopped, most stopped for lack of expertise, and very few of the firms that do use it have changed much about how the work gets done. At company level, diffusion has a plain meaning. A workflow that ran through five inboxes and two spreadsheets now runs differently: some steps are done by an agent, a person reviews the exceptions, someone owns the result, and there is a number from before and after. Then the next workflow. None of that needs a frontier lab. The best models are available to any company with an API key, and what decides the result is knowledge no lab can sell you: how your own orders, approvals and handoffs actually move, and which people need to trust the new way before it sticks. The export-control episode adds one rule I now apply in every build. The work gets redesigned around the process, and the model sits behind it as a part that can be swapped. If a vendor raises prices, falls behind or becomes unavailable in Europe for a quarter, the redesigned workflow stays and only the engine changes. That is the company-sized version of the insurance Benner and Hensing ask for. Draghi traced Europe’s last productivity gap to two failures, making digital technology and spreading it. The AI version has the same two halves. On AI, the first is mostly out of any single company’s hands. The second gets decided in thousands of companies that will never train a model and do not need to. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Fix a mistake once by turning every correction into a rule the agents read. · VANDAM > Turn every correction of an AI agent into a rule and a test with a named owner, and track the repeat-mistake rate. The part that survives a model change. URL: https://vandam.ai/notes/fix-a-mistake-once Contact: hendrik@vibelab.be # Fix a mistake once by turning every correction into a rule the agents read. Anthropic's data team took Claude from 21% to above 95% accuracy with written skills, and now turns every correction a colleague makes into a rule and a test for the person who owns that domain. I set out that loop for a mid-sized company, with where the rules live, who owns them, and the one number a COO should track: how often the same mistake comes back. Hendrik De Winne 26 October 2026 9 min read When Anthropic’s data team pointed Claude at its own warehouse without any written guidance, the agent answered business analytics questions correctly no more than **21%** of the time on the team’s evals. With a set of written skills added, accuracy climbed to **above 95%** in aggregate. Today **95%** of the company’s business analytics queries are automated, according to [How Anthropic enables self-service data analytics with Claude](https://claude.com/resources/articles/how-anthropic-enables-self-service-data-analytics-with-claude), a June 3 post by five members of that team. The part I keep rereading sits near the end of the post, in the plumbing. Every time a colleague corrects the agent in a thread, the team treats the correction as a candidate test. A scheduled agent scans the channels every few hours for correction language such as “that’s the wrong table,” drafts a one-line fix to the relevant reference document, and opens a pull request tagged to the person who owns that domain. The same corrections go into the test set. The authors call this “the part that closes the loop.” Most companies correct their agents all day and keep none of it. The fix happens in a reply, the reply scrolls away, and next week somebody makes the same correction again. The asset that compounds is the log of those corrections, the rules written from them and the tests built on top, owned by named people in the business. It is also the one part of an AI setup that survives the next change of model. - **2195%** Accuracy of Claude on analytics questions without and with written skills, per [Anthropic’s data team](https://claude.com/resources/articles/how-anthropic-enables-self-service-data-analytics-with-claude), June 2026. - **90%** Share of that team’s data-model changes that now update a skill file in the same change. - **20–50** Test cases drawn from real failures that [Anthropic’s engineers](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) call “a great start” for an eval suite. ## The jump came from written corrections The post credits the difference between 21% and 95% to written material: definitions of what each metric means, which tables to trust, which filters always apply, and what a finished analysis looks like. Much of it exists because someone got a wrong answer, said so, and the fix was written down where the agent reads it. The team treated this as engineering work. The skill files live in the same repository as the data models, so the change that alters a table is the same change that updates the document describing it. A domain owner cannot announce the agent to their colleagues until their slice of the test set passes a threshold, which the team set at about 90% at first. Every run is stored with the model version and skill version, so “did that change help?” becomes a query. One finding should interest anyone tempted to let the machine write its own rulebook. The team tried having a model generate metric definitions automatically from raw tables and query logs. The result “encoded the very ambiguities we were trying to eliminate” and did worse on their tests than a smaller set written by people. Their recommendation is to let Claude draft the documentation while “having a human own the definition.” ## A correction left in a chat gets made again The same habit shows up wherever agents do daily work. [Boris Cherny](https://x.com/bcherny/status/2007179840848597422), who leads Claude Code at Anthropic, described his team’s setup in January: one shared instruction file, checked into version control, that the whole team edits several times a week. “Anytime we see Claude do something incorrectly we add it to the CLAUDE.md, so Claude knows not to do it next time.” Other teams keep their own file, and “it is each team’s job to keep theirs up to date.” That is the whole mechanism, and it is cheap. It needs a habit and an owner. Without both, each correction costs a few minutes of a senior person’s time, the agent never learns from it, and the person correcting slowly concludes the agent is not worth the trouble. Writing the rule down is half the job. [Kasey Klimes](https://x.com/kaseyklimes/status/2093127970571317258), who builds decision software at [Primitive](https://www.getprimitive.ai) and so has a product in this space, put the gap well in August: “memory is half the equation. a company brain needs reliable enforcement of the decisions it has made.” A rule in a file is a request. A test that fails when the agent breaks the rule is enforcement. ## Tests turn a rule into something that outlives the model An eval is a test for an AI system: give it an input, check the output against an answer someone has agreed is right. Anthropic’s engineering team published a [guide to evals for agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) in January, and its advice for where to start is plain: look at the bug tracker and the support queue, because “converting user-reported failures into test cases ensures your suite reflects actual usage.” Tests that check whether the agent still handles everything it used to “should have a nearly 100% pass rate.” On who writes them: “The people closest to product requirements and users are best positioned to define success.” The guide also explains why this matters to a budget. “When more powerful models come out, teams without evals face weeks of testing while competitors with evals can quickly determine the model’s strengths, tune their prompts, and upgrade in days.” A suite built from your own corrections tells you within days whether a new or cheaper model is good enough. Satya Nadella turned the same idea into a test of ownership in a long post in June. > A company should be able to switch out a “generalist” model without losing the “company veteran” expertise built into their learning system. This is the key “test” of your control and sovereignty in the era ahead. > > Satya Nadella, Chairman and CEO, Microsoft, June 2026, [Source](https://x.com/satyanadella/status/2066182223213293753) In the same post he writes that “private evals should capture whether a model is actually improving against outcomes that matter to the business (not just external benchmarks!).” Nadella runs a company that sells the platform this loop would sit on, so read it as a vendor’s position as well as a strategy. [Garrett Lord](https://x.com/GarrettLord/status/2068754262440767500), CEO of [Handshake](https://joinhandshake.com), quoted him a week later and asked “how many business leaders can actually quantify how accurate their AI programs are?” Handshake sells evals, so that is a seller’s question, and still a fair one to put to your own team. I turn Nadella’s test into a budget rule. Go down the AI spend line by line and ask what would be left if the model changed tomorrow. Licenses, prompts tuned to one model’s quirks and fine-tunes tied to one model generation would mostly have to be redone, so I treat them as rent. The correction log, the rules written from it and the tests built on them carry over to whatever model comes next, so I treat them as the asset. In most companies the asset column is close to empty, and nobody is named as its owner. ## The method: from correction to rule to test This is the loop we set up, one process at a time. It needs a folder, a habit and a name to start. 1. **Capture every correction in one place.** When someone fixes an agent’s output, they log it in a shared list: what the agent produced, what it should have produced, and why. One line each. If corrections happen in chat, a weekly sweep by hand works until an agent can scan for them. 2. **Turn the correction into a rule.** The owner writes one plain sentence the agent can follow, such as “Freight from supplier X is always billed on the second invoice, so never flag it as missing on the first.” Each rule gets an ID, a date and the correction it came from. 3. **Store the rules where the agent reads them.** One plain text file per process, kept in a versioned folder the company controls, loaded by the agent before every run. The vendor’s settings screen is the wrong place, because the rules then leave with the vendor. 4. **Write the test from the corrected example.** The original input and the corrected output become a test case. Start with the 20 to 50 cases from last month’s corrections, and rerun the set on every change of model, rule or data source. 5. **Give each process a named owner in the business.** The controller owns the close rules and tests. The sales operations lead owns the quoting rules and tests. The owner approves new rules, removes stale ones and signs off before a new model goes live, the same way someone owns a general ledger account. The owner is the step companies skip, and it is the one that decides whether the rest holds. Only the person who knows why freight from supplier X is billed late can say whether a rule is right. That also gives auditors a clean trail of how the agent was supervised, which fits the controls I set out in [an earlier note on agents in the ledger](https://vandam.ai/notes/agents-and-internal-controls). ## Track one number: the repeat-mistake rate Each month, count the corrections logged for a process and mark the ones that repeat a mistake the agent has made before. The repeat-mistake rate is the share of corrections that are repeats. If it falls, the loop works. If it stays flat while the agent runs more jobs, people are correcting by hand and nothing is being kept. The repeats split in two, and each half points at a different fix. A repeat with no rule behind it is a capture failure: someone corrected the agent and nobody wrote it down. A repeat that breaks a rule already on file is an enforcement failure: the rule is unclear, the agent is not reading it, or it needs a test that fails loudly. I would put both counts in the monthly operations review next to the process’s volume and error figures. If you do not yet know what the process looked like before the agent, start with [the baseline](https://vandam.ai/notes/measure-before-you-automate) first. ## Rules go stale, so someone has to delete them A log that only grows becomes its own problem. Anthropic’s data team writes that it regularly prunes skill scaffolding “as models improve and previous failure modes no longer apply.” A rule written for last year’s model can block the better behavior of this year’s, and a rule written around a broken system keeps the workaround alive after the system is fixed. The owner’s job includes a quarterly pass to remove rules whose test now passes without them. The loop also misses some mistakes. The Anthropic post is candid that “the failure mode none of this fully catches is the silent one,” a wrong answer that looks plausible and is used without objection. Its mitigations include human sign-off on anything going to leadership. Corrections only exist where someone checks, so the processes with no reviewer will show a low repeat rate for the wrong reason. Recording why people made a decision, which I covered in [a note on what a company knows](https://vandam.ai/notes/context-compounds), is a separate habit. This loop works on the agent’s daily output. ## What a COO can do in the next 30 days Pick one process where an agent already produces work that people check, such as invoice coding, quote drafting or the first pass of the month-end reconciliation. Name one owner from the business side. Create the rules file and the correction log, and ask everyone who reviews the agent’s output to log each fix for four weeks. At the end of the month, have the owner turn the corrections into rules and the first 20 to 50 test cases, and compute the repeat-mistake rate for the month. Put that number on the agenda of the next operations review. Before the next AI contract renewal, add one clause: the rules, the correction log and the tests are stored in plain files the company owns and can take to another model. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # What Anthropic found when its agents outworked its people. · VANDAM > A close reading of Anthropic's report on its own jobs: agents at 6 to 1, attention as the bottleneck, fewer juniors, and lessons for a 300-person firm. URL: https://vandam.ai/notes/inside-anthropic Contact: hendrik@vibelab.be # What Anthropic found when its agents outworked its people. Anthropic measured its own company after agent hours passed human hours six to one, and the findings are about attention, management, hiring and people moving. I read the report as a story and draw out what it means for a company of 300. Hendrik De Winne 24 October 2026 8 min read By August 2026, the agents that Anthropic employees ran from their own desks worked six hours for every hour their people worked. That number sits in the first paragraph of [How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic), a report four researchers on the company’s economics team published on October 16. They used internal headcount, agent telemetry, calendars, Slack, hiring scorecards and 17 interviews, and turned all of it on their own employer. It is the first time I have seen a company this far into agents publish what happened to its own jobs, and the findings are mostly about organization. The model capabilities are background. What the report keeps returning to is attention, management, hiring and who moves where. Those are the questions every company I work with will face, a few years later and with less money to absorb the mistakes. - **6:1** Agent hours to human hours for employee-directed sessions at [Anthropic](https://www.anthropic.com/research/employment-at-anthropic), August 2026. - **80%** Share of interactive Claude session time spent idly waiting for human input. - **9%** Share of technical hires at junior level, down from 27%. - **22%** Share of interactive sessions spent on meta-work such as specs, reviews and agent dashboards, up from 10%. ## The hiring changed first The story starts in January 2026, when engineers inside Anthropic began talking about needing fewer juniors. The teams that leaned hardest on Claude found that writing code and finding bugs could increasingly be handed off. By April, engineering and research had deemphasized hiring for lower band technical roles. The share of junior employees in technical departments slipped from **21%** to **14%** since early 2025, and the share of junior technical hires fell faster, from **27%** to **9%**, according to [the report](https://www.anthropic.com/research/employment-at-anthropic). The company kept growing, so this is a story about mix. Technical departments went from 61% to 56% of headcount in a year while sales, finance and recruiting hired faster. Part of that is a scaling company adding business functions. Part of it is non-technical people building their own software: coding agent sessions for employees with no code commits at baseline went from 2 a month in August 2025 to **90** in May 2026. A sales rep built a custom dashboard and pipeline in three to four hours. Administrative roles, the ones that usually scale with headcount, barely grew at all. ## Then the bottleneck moved to the people Once agents were doing most of the hours, the scarce resource became a human paying attention. The report measures it directly: in interactive sessions, Claude spends around **80%** of its time waiting for a person to respond. Even for non-technical teams, reading everything their agents produced would take more than half the work week. One engineer described the job as facing “an army of Claudes, with every one asking for my attention.” Another reached for childcare. > When you are babysitting, every child brings to you concerns with an equal demand for your attention and the various [Claudes] also do that. > > An Anthropic engineer, Anthropic, Interviewed for the October 2026 report, [Source](https://www.anthropic.com/research/employment-at-anthropic) This is the part most AI business cases miss. They count the hours a model saves and assume the output lands somewhere useful. At Anthropic the output piles up in front of a person, and the person becomes the limit on how much of it turns into anything. ## So people started managing Claudes The response was a new layer of work about the work. Meta-work, meaning skills, plans, reviews for other agents, dashboards and documentation, grew from **10%** of the average employee’s interactive sessions in fall 2025 to **22%** in spring 2026. Written standing instructions for agents went from a few hundred in July 2025 to **8,500** a year later. One engineer set up an agent chief of staff to triage agent requests. Another wrote a performance review for their Claude team. Meetings went up too. The median employee spends 20% more time in meetings than a year ago, and technical individual contributors 16% more. The authors read it as the hard calls moving upward: when an agent chases a bug far enough, it hits organizational problems that need people in a room. And because agents only know what is written down, private decisions became a liability. One employee kept getting agent requests for t-shirts and mugs because their name was still on a months-old document about swag. A non-technical employee summed up where the job went: “that 20% that was strategic becomes 90% of the job.” ## Hiring now looks for judgment The hiring scorecards show the same shift. Mentions of a candidate’s track record with AI tools tripled for technical roles and quadrupled for non-technical roles between 2024 and 2026. For technical candidates, the biggest increases were in comments on autonomy and judgment, while domain expertise shrank as a share of the assessment. One manager put it plainly: “I expect models to continue to be jagged, so for my team I want them to be fungible across domains.” That matches what [Boris Cherny](https://x.com/bcherny/status/2071379474277613732), who leads Claude Code at Anthropic, wrote in June. He sees five archetypes on his team: the Prototyper who churns out ideas, the Builder who turns one into production, the Sweeper who simplifies and unships, the Grower who iterates toward product-market fit, and the Maintainer who keeps a mature system secure and fast. A healthy team mixes them depending on the product’s stage, and he notices they cut across job titles, with designers, engineers, PMs and data scientists in every category. > Maybe product roles of the future will look more like this, and less like the domain-specific roles of today? > > Boris Cherny, Head of Claude Code, Anthropic, June 2026, [Source](https://x.com/bcherny/status/2071379474277613732) ## People moved instead of leaving The last finding is the one I find most useful. Employees with at least a year at the company were **50%** more likely to move across departments each month in 2026 than in 2025. Whole teams moved. A product design team that wrote the greetings on Claude.ai built a context graph that let agents write that content, then rotated into writing prompts and model evals. Their manager’s line: “Writers who used language to drive human behavior are now using it to drive Claude behavior.” Anthropic is now standing up an early warning system that reads role-level demand and productivity against AI capabilities, so it can move people before their work disappears. The authors argue that committing to redeployment pushes a company to look for new tasks and new products, and helps it resist what economists call “so-so” automation: replacing workers with technology that barely beats them, which costs the displaced people a lot and earns the company almost nothing. They also credit broad stock ownership, and say plainly that many of their practices cannot be copied. They also name the weak spot in their own plan, the juniors who never got hired. Employment protections for incumbents will not benefit those not hired to begin with. Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory, economics team at [Anthropic](https://www.anthropic.com/research/employment-at-anthropic) There are caveats worth keeping in view. Anthropic sells the model, grew explosively in the same period, and tripled in size in a year. The productivity comparisons follow 192 engineers who adopted agents against 74 who did not, and the authors note that adoption was not random. Treat it as one company’s evidence, collected by people with every reason to look carefully. ## What I take from it for a 300-person company Everything below is my reading. Anthropic makes none of these claims about other companies. **Roles.** Most mid-sized companies still organize around functions, and the job descriptions describe execution. If execution is what agents absorb first, those descriptions will age fast. I would start sorting people by what Cherny’s list describes, how they work on a product or process at a given stage, and staff teams by stage. A plant running a mature ERP needs Maintainers and Sweepers. A new service line needs Prototypers and Builders, whatever their title says today. **Managing agents.** Somebody has to own the attention budget. In a company of 300, nobody yet asks how many hours of agent output land on each person’s desk every week, who reviews it, and which standing instructions are current. I would treat those instructions like any other operational document, with an owner and a review date, and I would push decisions into writing so the agents work from the same context as the people. The t-shirt story will happen everywhere. **HR.** The junior question is the hardest one. A company with ten junior analysts today may want three next year, and then has no one to promote into the judgment roles in five. I would decide on purpose how juniors still get time on the object-level work that builds judgment, and I would build internal mobility before the cuts come: a simple list of roles under pressure, read every quarter, and a path out of each one. **Finance.** I would stop approving business cases that only count headcount removed. That is the so-so automation the report warns about, and the gains tend to stay small. The cases I trust count agent hours, the human review time they consume, and the new work the freed people take on. If a case has no line for the third, it is a cost cut. The first step costs nothing. Pick one team that already uses agents heavily, measure for two weeks how much agent output each person has to read and approve, and put that number next to the hours the agents saved. That ratio tells you where your bottleneck is. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Before you automate anything, find out where the hours actually go. · VANDAM > Where the hours really go before AI: coordination, handoffs and exceptions. The evidence, and a two-week method to rank tasks by hours times feasibility. URL: https://vandam.ai/notes/map-the-work-first Contact: hendrik@vibelab.be # Before you automate anything, find out where the hours actually go. Most AI shortlists are built from the visible work, while the hours sit in coordination, waiting and exceptions that no process chart shows. The evidence from 1990 to 2026, and the two-week work map we use to rank tasks by hours times feasibility. Hendrik De Winne 25 October 2026 8 min read In the summer of 1990, Michael Hammer put a small aside in parentheses in his [Harvard Business Review](https://hbr.org/1990/07/reengineering-work-dont-automate-obliterate) article “Reengineering Work: Don’t Automate, Obliterate.” An insurer had estimated that while an application spent **22 days** in process, somebody actually worked on it for **17 minutes**. Everything else was waiting, passing and checking. I think about that ratio every time a leadership team shows me its AI shortlist. The list is almost always built from the visible work: the drafting, the data entry, the report someone spends Friday afternoon on. Those tasks are real and they are easy to demo. They are rarely where the hours go. Before we automate anything, we find out where the time actually sits, and most of the time it sits somewhere the shortlist never mentions. - **17min** Hands-on work in an insurance application that spent 22 days in process, per [HBR](https://hbr.org/1990/07/reengineering-work-dont-automate-obliterate), 1990. - **57%** Share of the average employee’s time spent in meetings, email and chat, per [Microsoft](https://www.microsoft.com/en-us/worklab/work-trend-index/will-ai-fix-work), 2023. - **900k** Distinct paths one multinational’s order-to-cash process took in a single year, per [Wil van der Aalst](https://www.vdaalst.com/publications/p1089.pdf), 2020. - **41%** Share of startup-to-task mappings that land on tasks workers don’t want automated, per [Stanford](https://arxiv.org/abs/2506.06576), 2025. ## The hours hide between the steps Hammer’s main case in that article was Mutual Benefit Life, then the eighteenth largest life carrier in the US. An application could pass through as many as **30** discrete steps, across **5** departments and **19** people. Typical turnaround ran from 5 to 25 days, “most of the time spent passing information from one department to the next.” Nobody inside any single department would have described their own step as slow. The delay lived in the handoffs, and handoffs belong to no one. The fix was a new job. MBL created a case manager who owned an application from receipt to issued policy, with specialists on call as advisers. Turnaround dropped to two to five days, 100 field office positions went away, and case managers handled more than twice the volume the old process could. The same article mentions an East Coast insurer whose regional offices sent a series of reports to the home office for years. Nobody in the field knew they were simply filed and never read. The title of the article is the instruction. Hammer argued that companies kept using computers to speed up processes designed before computers existed, and that the gains stayed small because the process itself was the problem. Ford’s accounts payable redesign, which I walked through in [an earlier note](https://vandam.ai/notes/ten-steps-one-is-ai), came from the same piece. Swap “computers” for “agents” and the argument needs no other edits. ## Most of the day goes to work about the work Thirty-five years later the waiting has moved into inboxes and calendars. Microsoft’s [2023 Work Trend Index](https://www.microsoft.com/en-us/worklab/work-trend-index/will-ai-fix-work), based on Microsoft 365 activity data, found the average employee spends **57%** of their time communicating in meetings, email and chat, and 43% creating in documents, spreadsheets and presentations. In the same report, 62% said they spend too much time searching for information. The June 2025 edition, [Breaking down the infinite workday](https://www.microsoft.com/en-us/worklab/work-trend-index/breaking-down-infinite-workday), got more granular. The average worker receives **117** emails and **153** Teams messages a day, and is interrupted by a meeting, email or notification every two minutes during core hours, about 275 times a day. Microsoft found that 57% of meetings were ad hoc calls with no calendar invite, which means they never show up in anyone’s picture of the week. Asana’s [Anatomy of Work Index 2021](https://asana.com/resources/anatomy-of-work-index), a survey of over 13,000 knowledge workers in seven countries, put a name on this: “work about work.” Respondents said they spend **60%** of their day on it, meaning chasing status, searching for information, switching between apps and sitting in meetings about the work, and only the rest on the skilled job they were hired for. Unnecessary meetings alone cost each of them 157 hours in the past year. These are vendor studies, and the Asana figures are self-reported, so I hold the exact percentages loosely. The direction matters more. When a COO pictures where the order desk spends its time, the picture is of people doing orders. The data says that more than half the day goes to the coordination around the orders, and none of that coordination appears on a process chart. ## The documented process covers a minority of real cases Process mining makes this visible with system data instead of surveys. The software reads timestamps from an ERP or ticketing system and reconstructs every path a case actually took. Wil van der Aalst, the RWTH Aachen professor who built the first process mining techniques, describes in a [2020 paper on shared services](https://www.vdaalst.com/publications/p1089.pdf) the order-to-cash process of a large multinational: about 30 million cases a year, over 60 possible activities, and more than **900,000** distinct process variants in a single year. The most common variant occurred over 3 million times, roughly one case in ten. His summary is that a small share of variants explains most cases, while the remaining 20% of cases account for 80% of the variants. Many of those rare paths “involve rework, passing the buck (leaving a difficult problem for someone else to deal with), communication errors, and repair actions.” He calls most deviations from the happy path operational friction. That has a direct consequence for AI. An agent built on the process as documented will handle the happy path well. The happy path is also the part that already runs fastest. The exceptions, where the hours pile up and customers wait, are exactly the cases the documentation leaves out. ## AI does best on short, checkable tasks The newest benchmarks measure AI at the level of a task, which is the level a work map needs. OpenAI’s [GDPval](https://arxiv.org/abs/2510.04374), published in October 2025, built **1,320** tasks from the actual work product of professionals in 44 occupations, people with an average of 14 years of experience. The tasks took an expert about 7 hours on average. On the public subset, expert graders rated **47.6%** of Claude Opus 4.1’s deliverables as better than or as good as the human one. Win rates were highest for tasks of 0 to 2 hours and declined steadily as tasks got longer. Anthropic’s [Economic Index report](https://www.anthropic.com/research/economic-index-primitives) from January 2026 reads real usage instead of a test set. Claude completes tasks that require a college degree **66%** of the time, and through the API it reaches 50% success on tasks that would take a person about 3.5 hours. When the authors adjusted their productivity estimate for those failure rates, it fell by about a third. Reliability, more than raw capability, decides how much of an estimated saving survives contact with real work. Put next to the coordination data, this sharpens the question. Feasibility is a property of a specific task with a specific input and a specific check at the end. Nobody can score “accounts receivable” for feasibility. Someone can score “match a remittance email to open invoices and flag the ones that don’t reconcile.” A map at that resolution is the only thing that lets you put hours and feasibility in the same row. ## The guess from the top usually misses The clearest evidence that leaders guess wrong comes from a Stanford team led by Yijia Shao and Diyi Yang, with Erik Brynjolfsson, in [Future of Work with AI Agents](https://arxiv.org/abs/2506.06576). They asked 1,500 workers across 104 occupations about 844 of their own tasks, and had AI experts rate what agents could do on each. Workers were positive about automation for **46.1%** of tasks, mostly to free time for higher-value work. Then the team mapped Y Combinator startups to the same tasks: **41.0%** of those mappings landed on tasks where workers did not want automation, whether or not the technology could do them. The ten occupations where workers most wanted automation accounted for only 1.26% of Claude.ai usage in the period they compared. Startup investment clustered around software development and business analysis, and many of the tasks workers wanted help with, and that agents could already do, were left largely unaddressed. That is the same error at market scale that I see inside single companies, where the AI roadmap is written by the people furthest from the work. ## A two-week work map This is the method we use. It costs two weeks and one person who can run interviews and a spreadsheet. **Scope one value stream.** Pick a flow with a clear start and end that crosses at least two teams: quote to order, order to cash, month-end close, claims intake, hire to first day. Twenty to eighty people is the right size. Name an owner who can change the process afterwards. **Interview eight to twelve people in days one to three.** Two front-line people from each role that touches the flow, the person everyone sends the strange cases to, the team that receives the output downstream, and whoever administers the main system. Managers come last, so their picture does not anchor the others. Ask each person to walk through the last three cases they handled, in order, with the screens open. Real recent cases surface the workarounds that a description of a typical day leaves out. **Log time for five working days, days three to eight.** Everyone in scope fills a simple sheet in 30-minute blocks, with one task name per block and one of four labels: core work, coordination (meetings, email and chat about the work), waiting or chasing, and rework or exceptions. Five days is enough to catch a weekly cycle and short enough that people keep filling it in. In parallel, pull 90 days of timestamps from the ERP, CRM or ticketing system and calculate, per case, elapsed time against touch time. You do not need process mining software for a first pass. A count of how many cases followed the standard sequence will do. **Score every task in days nine and ten.** Build one table with a row per task and four columns that matter. Hours per week across the team, from the logs. A feasibility score from 0 to 3, one point each for: the inputs are digital and reachable, the output can be checked by a person in minutes, and the task fits in under two hours of expert work or splits into pieces that do. A worker-preference flag from the interviews, following the Stanford finding. And Hammer’s question, asked of every row before any other: does this task need to exist at all? **Rank and decide.** Tasks that should not exist go on a separate list and get removed first, because deleting a report nobody reads beats automating it. Everything else is ranked by hours times feasibility. Our decision rule is simple: in this round, only the top three rows get an AI build, each with a baseline from the logs and an owner. A task that scores well on feasibility but sits low on hours waits, however good the demo looks. The first output is usually uncomfortable. The top of the hours column is coordination and exceptions, and some of the biggest fixes are a decision right, a shared field in the ERP or a meeting that stops. Those ship in week three, before any agent. If you run operations, pick the value stream this week, name its owner, and send the time-log sheet with a start date. Book the interviews with front-line staff before the managers. In two weeks you will have a ranked list with hours attached, and that list is what the AI budget should be approved against. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Measure the work before you let AI change it. · VANDAM > Why most firms cannot prove AI ROI, and a two to four week method for CFOs to baseline hours, cycle time, rework and cost per unit before AI. URL: https://vandam.ai/notes/measure-before-you-automate Contact: hendrik@vibelab.be # Measure the work before you let AI change it. Eight in ten survey respondents say AI makes them more productive, yet only about a third of companies can find it in EBIT, and few measured the work before it changed. Why the companies with credible AI numbers all name a starting point, and a three-week baseline method a CFO can run before the next AI purchase. Hendrik De Winne 25 October 2026 8 min read On 29 July, C.H. Robinson filed its second quarter results with the [SEC](https://www.sec.gov/Archives/edgar/data/0001043277/000104327726000027/ex991earningsreleaseq20630.htm). The freight broker credits what it calls its Lean AI strategy with productivity improvements of **over 60%** “since the end of 2022” in its two main businesses. The same document shows average headcount down from **12,858** to **11,471** in a year, North American volume up 1.5% while the industry’s shipment index fell 3.3%, and adjusted operating income up about 20%. The phrase I keep coming back to is the date. C.H. Robinson picked a starting line, kept counting from it, and now reports against it every quarter. Most companies I meet cannot do that for a single workflow, because nobody wrote down how long the work took, how often it went wrong, or what one finished unit cost before the AI arrived. Without that, a CFO is left defending AI spend with anecdotes and survey scores. - **37%** Respondents who attribute any EBIT impact to AI, flat on 2025, per [McKinsey](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai), 2026. - **25%** AI initiatives that delivered the expected ROI, according to 2,000 CEOs surveyed by [IBM](https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles), 2025. - **8%** Organizations reporting established AI ROI, while 95% have an AI strategy, per [KPMG](https://assets-launch.kpmg.com/content/dam/kpmgsites/xx/pdf/2026/04/global-ai-pulse-exec-summary.pdf.coredownload.inline.pdf), Q1 2026. - **6%** Executives who saw payback on a typical AI use case within a year, per [Deloitte](https://www.deloitte.com/global/en/issues/ai/ai-roi-the-paradox-of-rising-investment-and-elusive-returns.html), 2025. ## People feel faster while the P&L stays flat The clearest version of the gap comes from [The state of AI in 2026](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai), which McKinsey published in August from a survey of 1,719 people in May and June. Eight in ten respondents say AI has improved their own productivity. Only **37%** say it has contributed anything to their organization’s EBIT, essentially the same share as a year earlier, even though more companies are now scaling AI. The high performers, those who attribute at least 5% of EBIT to AI, are stuck at about **6%** of respondents. The other big surveys land in the same place. [KPMG’s Global AI Pulse](https://assets-launch.kpmg.com/content/dam/kpmgsites/xx/pdf/2026/04/global-ai-pulse-exec-summary.pdf.coredownload.inline.pdf), 2,110 executives in 20 countries, found that 64% report meaningful business value and only 8% report established return on investment. [IBM’s 2025 CEO study](https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles) holds the oddest pair of numbers: 68% of CEOs say their organization has clear metrics to measure innovation ROI, and the same CEOs say only 25% of their AI initiatives delivered the return they expected. Having a metric on a slide and having a number from before the change are two different things. The difference between the 6% and everyone else is partly about measurement. McKinsey reports that high performers are twice as likely as others to have defined processes to measure the impact of their AI initiatives, and far more likely to redesign workflows instead of adding AI to existing ones. The second habit depends on the first. You cannot tell whether a redesigned workflow beats the old one if nobody recorded the old one. ## Without a before, the after cannot be audited Deloitte asked 1,854 executives in Europe and the Middle East about AI returns in the summer of 2025. Most said a typical use case reached satisfactory ROI in two to four years, against the seven to twelve months they expect from other technology. The [report](https://www.deloitte.com/global/en/issues/ai/ai-roi-the-paradox-of-rising-investment-and-elusive-returns.html) names the measurement problem directly: “fragmented systems and siloed platforms make it challenging to track before-and-after impact.” Executives added that AI usually arrives alongside data cleanup, team changes and process fixes, which makes its share of any gain hard to isolate. That second point is the one finance teams underrate. When a company rolls out an AI tool, cleans its vendor master and reorganizes the AP team in the same quarter, every one of those projects will claim the improvement. A baseline taken before any of them started, plus a team that changes later, is the only way to split the credit with a straight face. Gartner made the same argument to CFOs in May, using a survey of 204 finance leaders in which 63% said AI implementation in 2025 was slower than expected. > They must not mistake activity for impact. > > Marco Steecker, Director Analyst, Finance practice, Gartner, May 2026, [Source](https://www.gartner.com/en/newsroom/press-releases/2026-05-28-gartner-says-cfos-must-stop-mistaking-finance-ai-deployment-for-value-creation) He was talking about counts of pilots, tools rolled out and use cases in production. Those are the numbers most AI dashboards report, because they are easy to collect after the fact. Hours, errors and unit cost are hard to reconstruct after the fact, so they have to be collected first. ## The credible reports all name a starting point The public examples I trust share one feature: a stated before. C.H. Robinson fixes its start at the end of 2022 and puts the productivity claim in the same filing as headcount, volume against the market and operating margin, so an analyst can check one against the others. Its CEO, Dave Bozeman, also adds that the strategy “needs to result in better service,” and cites customer satisfaction scores next to the productivity number. The release does not spell out exactly how productivity is calculated, which is the one thing I would want added. [Klarna’s 2024 announcement](https://www.klarna.com/international/press/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month/) of its customer service assistant is older and widely quoted, and it reads like a baseline report. Resolution time fell from **11 minutes** to under 2. Repeat inquiries, the support version of rework, dropped **25%**. Customer satisfaction was on par with human agents. The $40 million profit figure was labeled an estimate for the year ahead, which is honest. Each claim compares the new state to a measured old one, and one of them measures quality alongside speed. The research version is a study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, first released in 2023 as an [NBER working paper](https://www.nber.org/papers/w31161). They followed 5,179 support agents as an AI assistant was rolled out to different groups at different times. Because they had each agent’s output before access and a group that had not yet received it, they could say with some confidence that issues resolved per hour rose **14%** on average and **34%** for novice and low-skilled agents, with little effect for the most experienced. A CFO does not need an economist to borrow that design: measure first, roll out in waves, and compare. ## Four numbers per workflow are enough A baseline does not need a data warehouse. For each workflow you plan to change, I would record four numbers, all tied to one clearly defined finished unit: a paid invoice, a closed ticket, a signed contract, a shipped order, a closed month. **Hours per unit.** Total human hours spent on the workflow in the period, divided by units finished. Include review, chasing and handoffs, which are usually where the time goes. **Cycle time.** Elapsed time from the trigger to the finished unit, taken from system timestamps where they exist. Report the median and the slowest tenth, since the slow tail is what customers and auditors notice. **Error and rework rate.** The share of units that come back: reopened tickets, credit memos, correcting journal entries, contract redlines after signature. This is the number AI can quietly make worse while hours improve, so it belongs in the baseline from day one. **Cost per finished unit.** Fully loaded labor for those hours, plus software, outside services and the cost of rework, divided by units. Once AI is in the workflow, add token and license costs and the human time spent checking its output. I made the case for pricing AI this way in [an earlier note](https://vandam.ai/notes/cost-of-the-i); the baseline is what gives that unit cost something to be compared with. ## A baseline a CFO can run in three weeks This is the method we use. It fits inside two to four weeks, depending on how many workflows you take on, and it needs one analyst from FP&A, the process owner and a few hours from the people who do the work. **Week one: choose and define.** Pick the three to five workflows where AI money is about to be spent. For each one, write a single sentence defining the finished unit and when it counts as done. Agree on the period you will measure, ideally a normal month with no close or peak season distortion, or adjust for it. Name the controller who will sign the result. **Week two: collect.** Pull counts and timestamps from the systems that already hold them: ERP, ticketing, CRM, document management. For hours, ask the team to log time on the workflow in 15-minute blocks for ten working days, or use work sampling, where people record what they are doing at a few random moments each day. Pull rework from the system records listed above and ask the team which kinds of rework the system misses. **Week three: price and freeze.** Turn hours into cost with the loaded rates finance already uses. Calculate the four numbers per workflow and put them on one page with the definitions, data sources and known gaps. The process owner and the controller sign it, and the page is filed where the audit trail lives. If possible, hold one team, site or region back for the first wave of the rollout, so you have a group to compare against. **Week four, if you need it: set the rules.** Agree in writing how savings will be counted. My rule is that a saving counts only when it shows up in one of three places: fewer hours bought (overtime, contractors, open requisitions left unfilled), more finished units for the same cost, or lower rework cost. Set review dates at 90 and 180 days, measured exactly the way the baseline was. The decision rule that follows is simple. No AI business case gets approved without a signed baseline for the workflow it targets. That one rule changes the conversation with vendors, internal champions and the board, because every claim afterward is a comparison against a number someone signed before anything changed. ## What to do this quarter For the CFO: pick the one workflow where the next AI purchase is most likely, run the three-week baseline before the contract is signed, and attach the one-page result to the purchase request. Then ask for the same page from every team that already rolled out AI without one, and accept that their first honest measurement starts today. For the CEO: ask for AI results in the board pack in the same four numbers, against a stated start date, the way C.H. Robinson reports them to its shareholders. If a team cannot show a before, give it three weeks to get one before you approve its next request. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Most of your AI list needs software and only a few items need an agent. · VANDAM > Most AI requests are rules for software. Agent reliability and cost data, four tests, and ten mid-market requests sorted into software, agent or person. URL: https://vandam.ai/notes/most-of-it-needs-software Contact: hendrik@vibelab.be # Most of your AI list needs software and only a few items need an agent. Agents still finish a minority of office tasks and give different answers on repeated runs, while a written rule does the same thing every time for almost nothing. I set out four questions that sort every request into software, agent or person, and apply them to ten typical ones. Hendrik De Winne 26 October 2026 9 min read Researchers at Carnegie Mellon built a mock software company, with a chat system, a file server and simulated colleagues, and gave AI agents 175 ordinary office tasks to finish on their own. In the latest version of [TheAgentCompany](https://arxiv.org/abs/2412.14161), revised in September 2025, the best agent completed **30.3%** of them, at an average model cost of **$4.20** per task. The weakest categories were finance, administration and data science, “with many LLMs completing none of the tasks successfully.” Building spreadsheets, collecting information from several people and reading scanned documents are easy for a person and hard for an agent, the authors note. Those are the tasks that fill most mid-market AI lists. Read such a list line by line and most of the items need no agent at all. Routing an invoice to the right approver, pushing a closed deal from the CRM into billing, checking that a new hire’s forms are complete: each is a rule someone can write down. A rule in software runs the same way ten thousand times, costs a fraction of a cent per run, is tested once and can be shown to an auditor. An agent doing the same job costs more per run and may give a different answer the second time. So the most valuable AI decision a COO or CFO can make early is a sort. Every item goes to software, to an agent, or to a person. Then a second filter: an item only gets automated if the work repeats and something other than a person can mark the output wrong. - **30%** Office tasks the best agent finished on its own in [TheAgentCompany](https://arxiv.org/abs/2412.14161), Carnegie Mellon, 2025. - **30–75%** Range of outcome consistency on repeated runs of the same task across 14 models, per [Princeton](https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai), 2026. - **15×** Tokens a multi-agent system uses compared with a chat, per [Anthropic](https://www.anthropic.com/engineering/multi-agent-research-system), 2025. ## Back-office work is where agents score worst The benchmark is useful because it looks like a real workplace. Some admin and HR tasks contain deliberate traps, and one finance task asks the agent to complete a section of IRS Form 6765, the research credit, from company financials on the file server, and to contact the finance director when something is unclear. Coding came out best, partly because models train on public code, the authors suggest, while financial data sits privately inside companies. The models in the paper are a year old, and newer ones score higher. The cost line matters more. At $4.20 per attempt and seven attempts in ten failing, the best agent’s cost per completed task works out to nearly $14. For work a rule could have done, all of it is overhead. ## The same agent gives a different answer the second time The stronger evidence for a CFO is about consistency. In February 2026, Stephan Rabanser, Sayash Kapoor, Arvind Narayanan and colleagues at Princeton published [Towards a Science of AI Agent Reliability](https://arxiv.org/abs/2602.16666). They ran models from OpenAI, Google and Anthropic released over 18 months through two benchmarks, five times per task, with paraphrased instructions. In [their summary](https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai), they write that “nearly two years of rapid capability progress have produced only modest reliability gains.” Many models gave inconsistent answers, “with outcome consistency scores ranging from 30% to 75% across the board,” and performance dropped substantially when the same instruction was reworded. One detail belongs in every finance team’s risk register. On τ-bench, a benchmark where agents serve simulated customers under a company policy, the paper finds that financial accuracy, meaning incorrect charges and refunds, “remains the most common failure mode across all models.” τ-bench itself, published in 2024 by [researchers at Sierra](https://arxiv.org/abs/2406.12045), introduced the measure that matters here: the chance an agent solves the same task correctly every time across eight tries. The best agent then solved about 61% of retail tasks on a single attempt and fewer than **25%** on all eight. The Princeton authors argue that agents need three to five nines, “99.9% to 99.999% accuracy,” before reliability stops being an issue, and “we don’t think LLM-based agents are on track to reach such a threshold.” A tested three-way match rule gives the same answer on every run. ## An agent doing a rules job costs more and proves less Anthropic, which sells agents, says the same in [its guide to building them](https://www.anthropic.com/engineering/building-effective-agents). It recommends “finding the simplest solution possible, and only increasing complexity when needed,” and notes that fixed workflows “offer predictability and consistency for well-defined tasks,” while “agentic systems often trade latency and cost for better task performance.” In its post on its own [research system](https://www.anthropic.com/engineering/multi-agent-research-system), the company reports that agents “typically use about 4× more tokens than chat interactions” and multi-agent systems about 15 times more. The academic version is older and sharper. In [AI Agents That Matter](https://arxiv.org/abs/2407.01502) (2024), Kapoor, Narayanan and colleagues showed that simple baselines beat several elaborate agent designs on a coding benchmark while costing much less, and that for similar accuracy “the cost can differ by almost two orders of magnitude.” [Annelies Gamble](https://x.com/AnneliesGamble/status/2069814714214023186) of Zetta Venture Partners, whose three-way split of work I used in [an earlier note](https://vandam.ai/notes/ai-needs-hr-and-finance), puts the test plainly in the same essay: “If the inputs are structured and the desired behavior can be specified, you probably do not need an agent. You need software.” Her examples are routing an invoice, syncing a CRM field and checking whether required documents are present. > Knowing what not to agentify is the first question to answer when it comes to enterprise deployments – because not every workflow deserves an agent. > > Annelies Gamble, Partner, Zetta Venture Partners, June 2026, [Source](https://x.com/AnneliesGamble/status/2069814714214023186) ## Automate only what repeats and what a machine can mark wrong The second filter comes from software engineering, where agents now run in loops that retry until the work passes. A June 2026 post by [Codez](https://x.com/0xCodez/status/2064374643729773029), an AI content creator summarizing an analysis by the AI newsletter AlphaSignal, lists four conditions for a loop to pay back. Two of them translate directly to operations: “The task repeats. A loop amortizes its setup across many runs.” And: “Verification is automated. The loop needs something that can fail the work without you in the room.” Without an automatic check, every agent run ends with a person reading the output, and the saving turns into a new review job. Anthropic measured that effect on its own staff, which I covered in [the note on its workforce report](https://vandam.ai/notes/inside-anthropic). Finance and operations already own many checks: the bank reconciliation, the match tolerance, the trial balance that must net to zero, the valid GL account list. Where such a check exists, an agent’s output can be failed and retried without anyone in the room. Where none exists, the honest cost of the agent includes the reviewer’s hours. ## Four questions sort every item on the list This is the rule we use. Take each item in this order and stop at the first answer. 1. **Can someone write the rule?** If the inputs come from a system or a form and a new employee could follow the steps from a one-page checklist, it goes to software: a workflow tool, an ERP or CRM rule, a script. Exceptions the rule cannot handle go to a queue. 2. **Who answers for the outcome?** If a mistake would cost a key customer relationship, breach a legal duty, affect someone’s job or move money above an approval threshold, a named person decides. AI can assemble the file, and the person signs. The same applies when nobody can yet describe what a good result looks like. 3. **Does it repeat?** My rule is at least weekly and at least a few dozen times a month. Below that, the setup, testing and upkeep of an agent will not pay back, and a person with a good chat assistant is the cheaper answer. 4. **Can a machine mark it wrong?** If an existing control, a reference total or a second system can fail the output, it is an agent candidate. If only a person can judge it, either budget the review hours into the business case or leave it manual for now. ## Ten typical requests, sorted A list of the kind a 400-person distributor might hand its AI team, run through the four questions: 1. **Route vendor invoices to the right approver.** Software. Cost center and amount decide the approver; the ERP’s approval workflow already does this. 2. **Match invoices to purchase orders and receipts.** Software. A tolerance rule clears the bulk automatically. 3. **Push closed-won deals from the CRM into billing.** Software. A field mapping and an integration, with failed records sent to a queue. 4. **Check that every new hire’s onboarding forms are complete.** Software. A checklist in the HR system with reminders. 5. **Build the weekly cash report from bank balances and receivables aging.** Software. A scheduled report; both sources are structured. 6. **Answer customer emails asking where an order is.** Agent. The wording varies and the agent has to find the order, but the answer can be checked against the shipping system before it goes out. 7. **Resolve the invoice match exceptions.** Agent. Gathering the PO history, receiving notes and supplier emails takes judgment, and the proposed resolution must again pass the tolerance check before a clerk approves the posting. 8. **Compare inbound supplier contracts with the standard terms.** Agent, with counsel signing off. Formats vary, the clause library is the check, and a lawyer decides on any deviation it flags. 9. **Set the credit limit for a large new customer.** Person. An agent can assemble the financials and payment history; the controller or CFO signs, because the exposure is theirs. 10. **Write each employee’s performance review.** Person. The trust between a manager and a report is the work, and no machine can mark the review wrong. Five of ten go to software, three to an agent, two to a person. The proportions will differ by company, but a large share of back-office requests are rules nobody has automated yet. Check whether some of those five are features the company already pays for in its ERP or HR system. ## Every agent still needs someone who can build and fix it The sort also changes who you need. Visual agent builders are sold as tools for operations staff. [Noah Epstein](https://x.com/NoahEpstein_/status/1978436984591745202), who promotes a competing tool in the same post, wrote after testing the main ones in October 2025: “They’re ALL ‘no-code for people who can code.’” His list of what each still requires reads like a job description: API authentication, webhook logic, error handling, data transformation and conditional branching. Treat it as one builder’s claim, made while selling an alternative. The software bucket needs a systems analyst who knows the ERP. The agent bucket needs that person plus an owner for the checks and test cases who reads the failures every week. Whatever ends up in an agent also enters your control environment, which I described in [the note on agents and the ledger](https://vandam.ai/notes/agents-and-internal-controls). A CFO should budget for those people as a line in the business case, next to the licenses. ## What to do with your list this month Put the current AI list in one spreadsheet with four columns: the rule in one sentence (or “cannot write it”), the person who answers for the outcome, the monthly volume, and the check that can fail the output. Fill it in a 90-minute session with the process owners; most items sort themselves once someone tries to write the rule. Fund the software rows first. They are the cheapest, they can be tested and audited, and they leave cleaner data for the agents that come later. For every agent row, the CFO should ask for three things before approving it: why a rule cannot do the job, which existing check will mark its output wrong, and the cost per completed task including review hours, the measure I set out in [the note on pricing AI per finished task](https://vandam.ai/notes/cost-of-the-i). Rows with a named owner and no check stay with that owner. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # The new org chart has written rules. · VANDAM > Block, Coinbase and Fiverr rewrote their org charts around AI in public. The five rules they share, what the layoffs hide, and what Klarna learned. URL: https://vandam.ai/notes/org-chart-redrawn Contact: hendrik@vibelab.be # The new org chart has written rules. Block, Coinbase and Fiverr published how they are reorganizing around AI. The rules are concrete, most of them came with layoffs, and the people side decides whether they work. Hendrik De Winne 24 October 2026 8 min read In the mid-1850s, Daniel McCallum was general superintendent of the New York and Erie Railroad, a system of more than 500 miles worked by thousands of people. The informal management that held smaller railroads together had stopped working, and trains were colliding. McCallum drew a chart of who reported to whom. That chart became the blueprint for the modern corporation, and most companies still run on a version of it. I took that story from an essay Block’s co-founder and CEO [Jack Dorsey](https://x.com/jack/status/2039003879841362278) published in March 2026 with Roelof Botha, a partner at [Sequoia](https://block.xyz/inside/from-hierarchy-to-intelligence) and Block’s lead independent director, called “From Hierarchy to Intelligence.” It goes back further than the railroad. A Roman legion was built from tents of eight soldiers under a decanus, ten tents to a century of eighty, six centuries to a cohort, ten cohorts to a legion of roughly 5,000. The authors call that nesting “an information routing protocol built around a simple human limitation: a leader can effectively manage somewhere between three and eight people.” The Prussian General Staff added officers whose job was to plan, process information and coordinate, which the essay calls “middle management before the term existed.” West Point engineers carried that thinking into the railroads, and McCallum put it on paper. The point of the history is what the layers were for. For two thousand years, every layer of management existed to carry information up, carry decisions down and keep units aligned, because a person can only keep track of a handful of other people. Most of what a manager does in a large company today is still that routing work. When software can hold a running picture of what is being built, what is blocked and where the money goes, a large share of that work moves into the system. Block’s essay says exactly that: “The question was never whether you needed layers. The question was whether humans were the only option for what those layers do.” That is a big claim from a CEO with an interest in it. What makes it worth reading is that several CEOs have now written down, in public, what they are actually changing. ## Four companies wrote down the new rules Fiverr went first. In September 2025 [Micha Kaufman](https://x.com/michakaufman/status/1967624550020985069), Fiverr’s founder and CEO, sent his staff a memo titled “Fiverr is going back to startup mode.” > Today, we are launching such a transformation for Fiverr, to turn Fiverr into an AI-first company that's leaner, faster, with a modern AI-focused tech infrastructure, a smaller team, each with substantially greater productivity, and far fewer management layers. > > Micha Kaufman, Founder and CEO, Fiverr, September 2025, [Source](https://x.com/michakaufman/status/1967624550020985069) In February 2026 [Block](https://x.com/jack/status/2027129697092731343) followed with a note to the whole company about “smaller and flatter teams.” A month later came the essay, which spells out the structure: a company world model built from the artifacts of a remote-first company, an intelligence layer on top, and people at the edge in three roles. In May [Brian Armstrong](https://x.com/brian_armstrong/status/2051616759145185723), the co-founder and CEO of [Coinbase](https://x.com/brian_armstrong/status/2051616759145185723), emailed everyone there that the company would be rebuilt “as an intelligence, with humans around the edge aligning it.” In June [Boris Cherny](https://x.com/bcherny/status/2071379474277613732), who leads Claude Code at Anthropic, described the roles he sees on his own team. Read side by side, the documents agree on more than I expected. Five rules come up again and again. 1. **Fewer layers, wider spans.** Coinbase is moving to “5 layers max below CEO/COO,” with leaders carrying “as many as 15+ direct reports.” Armstrong’s reason: “Layers slow things down and create coordination tax.” Fiverr asks for “far fewer management layers.” 2. **Every manager also builds.** Coinbase: “No pure managers: Every leader at Coinbase must also be a strong and active individual contributor.” Block’s version is the player-coach, who “still writes code or builds models or designs interfaces” and also develops the people around them. 3. **One person owns each problem.** Block’s Directly Responsible Individual “might own the problem of merchant churn in a specific segment for 90 days, with full authority to pull resources” from other teams. 4. **Roles stop following job titles.** Coinbase is testing “one person teams” with “engineers, designers, and product managers all in one role.” Cherny sees five archetypes, “Prototyper, Builder, Sweeper, Grower, Maintainer,” and notes they “are not really tied to job function.” 5. **The system carries the context.** In Block’s model, the world model replaces “the information that used to flow through layers of management,” so people at the edge can act without waiting for it to travel up and down. These are concrete. You can hold a current org chart next to them and count. How many layers sit between the person talking to the customer and the person who can say yes? How many managers spend their week relaying status? Who owns the churn problem, by name, for the next quarter? ## Most of these memos were also layoffs Each of these documents is also a goodbye letter, and I do not want to read the org design and skip that part. - **4,000+** People asked to leave or entering consultation at [Block](https://x.com/jack/status/2027129697092731343), from over 10,000 to just under 6,000. - **~14%** Of [Coinbase](https://x.com/brian_armstrong/status/2051616759145185723), cut in May 2026. - **~250** Team members leaving [Fiverr](https://x.com/michakaufman/status/1967624550020985069) in September 2025. The reasons differ, and the differences matter. Dorsey wrote that Block was “not making this decision because we’re in trouble. our business is strong.” Armstrong named two forces, and the first was the market: “we’re currently in a down market and need to adjust our cost structure now.” AI was the second. Anyone citing Coinbase as proof that AI flattens companies should say that the cut also came in a crypto downturn. The way the cuts were handled differs too. Dorsey kept Slack and email open so people could say goodbye, and wrote: “i’d rather it feel awkward and human than efficient and cold.” Kaufman paired the reduction with “upskilling our existing team.” Even the authors of the essay admit they have not finished the job yet. > Block is in the early stages of this transition. It will be a difficult one, and parts of it will likely break before they work. > > Jack Dorsey and Roelof Botha, CEO and lead independent director, Block, March 2026, [Source](https://block.xyz/inside/from-hierarchy-to-intelligence) ## The cautionary case is a year older In 2024 Klarna said its customer service chatbot did the work of **700** human agents. In May 2025 its CEO, [Sebastian Siemiatkowski](https://fortune.com/2025/05/09/klarna-ai-humans-return-on-investment/), told Bloomberg the company was recruiting people again so customers always have a human to talk to. > As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality. > > Sebastian Siemiatkowski, CEO, Klarna, May 2025, [Source](https://fortune.com/2025/05/09/klarna-ai-humans-return-on-investment/) This was a partial reversal. A Klarna spokesperson told Fortune the company still was not replacing people who left, apart from freelance support agents, and was “very much still AI-first.” The lesson I take from it is narrow and useful: when cost is the only thing you measure, quality is what you lose, and customers notice before the dashboard does. [Anthropic](https://www.anthropic.com/research/employment-at-anthropic) made the same warning in more formal terms this month, in its study of its own workforce. Agent hours there now outnumber human hours **6 to 1**, and the report still makes the case for moving people into new work. Employers replace human workers, imposing substantial social costs in lost income for displaced workers, while delivering minimal productivity benefits. (Anthropic on “so-so” automation, [How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic)) The report describes what the alternative looks like inside one company. Employees who had been there at least a year were **50%** more likely to move across departments each month in 2026 than in 2025, and whole teams were reallocated after handing their work to agents. Prioritizing reallocation, the authors write, “forces AI-adopting companies to consider ways that AI can create new tasks for new products and new markets.” ## Redesign should create capacity first My position is simple. The five rules above are good rules, and I would put versions of them in front of any leadership team. Fewer layers, builders who manage, one owner per problem and shared context make a company faster. What I would change is the order of operations. When a team stops spending its week routing information, it gets hours back. The first decision is what to do with those hours: the backlog nobody had time for, the market you never entered, the customers who waited too long for an answer. Cutting is the last option on that list, and if it comes first, you find out later which knowledge walked out with the people who left. Dorsey said as much about his own review: “i accept that we may have gotten some of them wrong.” The people side decides whether any of this works. Concretely, that means three things before the new chart goes live. Each new role gets written down, with what the person owns, what they decide and what they hand to agents, because a player-coach without a clear brief drifts back into managing. The people affected get consulted before the design is final, since they know where the real work happens and the chart usually does not. And the training starts before the switch, so that the people expected to run agents have done it on their own work first. The CEOs in this note wrote their rules down and published them, which makes them easy to argue with. If you are redrawing your own chart, I would do the same: write the rules, show them to the people they affect, and change them when they say something you missed. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Who owns what your AI learns. · VANDAM > Most companies rent their AI. Why YC asks for startups that help companies own their intelligence, and the four things a mid-sized company can own this year. URL: https://vandam.ai/notes/own-your-intelligence Contact: hendrik@vibelab.be # Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. Hendrik De Winne 29 October 2026 6 min read In its request list for the Winter 2027 batch, [Y Combinator](https://www.ycombinator.com/rfs#own-your-intelligence) asks for startups that help companies own their intelligence. The essay, by partner Jon Xu, opens with a sentence I would put on the wall of every AI steering committee: “Today, nearly every company building with AI rents its intelligence. You send tokens to a frontier lab, get answers back, and pay the bill.” Xu’s argument is that the trade made sense when the best models were years ahead of everything else. That gap has closed. “Open-weight models are now competitive with the top frontier models on meaningful benchmarks,” he writes, and companies “have become more aware of their dependency on model rental.” His prediction is that many will keep renting frontier models for their hardest problems and also “own specialized models that encode how their business works.” I agree with the direction, and I think most of the companies I work with should start somewhere smaller. Before a mid-sized company owns a model, it can own everything the model works with. That part is available today, costs little, and decides how much the company keeps when the model changes. ## Renting has a cost that is easy to miss The obvious cost of renting is the invoice. The less obvious one was spelled out in June by an engineer who posts as [dom](https://x.com/ryiacy/status/2066260212772679864). His reading of the market: the large labs sell forward-deployed engineers to big companies, build their workflows on proprietary models, and then use “the traces and context from this to build RL envs to improve the models.” In his words, “instead of buying this data they’re getting paid gigantic consulting fees to extract it.” That is one person’s interpretation, and the labs would describe their services differently. The structural point holds either way. When the knowledge of how your business works is written into someone else’s system, the learning curve belongs to them. dom ends with the choice he sees: companies either keep paying for that arrangement or “find a way to build and own their own model ecosystem,” which he describes as “some combination of AI models, evals, RL envs, and workflows.” The same theme runs through what investors and operators wrote this year. [Garry Tan](https://x.com/garrytan/status/2061076595474121073), who runs Y Combinator, put the bottleneck plainly in May: “The models are smart enough already. What is missing is the company-specific context locked in senior people heads.” [Nikita Bier](https://x.com/nikitabier/status/2109497987055640958) listed proprietary data first among the few moats left. ## Four things a company can own this year In October [Greg Isenberg](https://x.com/gregisenberg/status/2110414696037716372) came back from two days of dinners with founders and executives in San Francisco. The best teams, he wrote, “treat the model as swappable.” He also gave the cleanest breakdown I have seen of what sits around the model in a working agent: “the connectors for that industry’s data, evals written by people who know the work, rules for what the agent can and can’t touch, and a clean screen for a human to review.” Translated to a company of 50 or 300 people, that gives four assets, none of which needs a model of your own. 1. **Context.** The documents, decisions and exceptions an agent reads, kept as files in your own storage, with an owner for each part. This is what [compounds](https://vandam.ai/notes/context-compounds) while models come and go. 2. **Evals.** A set of real cases from your work, each with what a good answer contains, written with the people who do the job. [Garrett Lord](https://x.com/GarrettLord/status/2068754262440767500), CEO of [Handshake](https://joinhandshake.com/), which sells evals to labs and large companies, put it in one line: “In the world of agents, performance management is evals.” 3. **Rules.** Instructions, permissions and review steps, written in plain files that a person owns and dates, so they can be read, moved and corrected. 4. **Connections.** Access to your systems through an open standard such as MCP, so a new model plugs into the same doors as the old one. With these four in your hands, a new model becomes a test you run, with your evals telling you whether it is better at your work. Without them, every switch is a rebuild, and every vendor contract is harder to leave than it looks. If a new model drops next month, does your company get better or worse? (A question asked at one of those dinners, reported by Greg Isenberg, October 2026) ## Where owning the model comes in later The strongest objection to all of this is that frontier models keep improving and most companies cannot match them. I agree, and so does Xu: the hardest problems will keep going to rented frontier models. Lord makes the same split for an insurer, which “should probably buy a coding agent off the shelf and pay for frontier tokens, while also building proprietary agents that encode its unique underwriting decisions as a sovereign IP asset.” So the practical order is context, evals, rules and connections first, and a model of your own only when those exist. Xu describes the missing work in the same terms: someone has to “turn workflows into evals, expertise into training data, production failures into improvements.” A company that already has its evals and its curated context has done most of that work. If a regulator, a customer or its own data rules later require an open-weight model on its own servers, it can run that model against the same tests and see whether it is good enough. For transparency: VibeLab, the team behind VANDAM, is a Claude implementation partner, and for new builds we usually suggest Claude first. That is a choice about today’s best fit. The context, evals, rules and connections we build stay in the client’s systems and work with whichever model the client picks next. Our [models page](https://vandam.ai/models) lists the ones we work with. A quick test for your own company: pick one AI workflow that runs today and find its instructions, the cases it was tested on and the list of systems it can reach. If all three sit in places you control and can read, you own most of what matters. If they sit inside a vendor’s dashboard or one engineer’s laptop, that is the first thing to move. ## More notes 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. ## Fix a mistake once by turning every correction into a rule the agents read. Anthropic's data team took Claude from 21% to above 95% accuracy with written skills, and now turns every correction a colleague makes into a rule and a test for the person who owns that domain. I set out that loop for a mid-sized company, with where the rules live, who owns them, and the one number a COO should track: how often the same mistake comes back. --- # Plan for the week your AI vendor goes away, before someone else picks the date. · VANDAM > Treat your AI model provider as a single-source supplier: a tested second provider, owned prompts and evals, and API terms in every renewal. URL: https://vandam.ai/notes/plan-for-your-ai-vendor-leaving Contact: hendrik@vibelab.be # Plan for the week your AI vendor goes away, before someone else picks the date. In February 2026 defense contractors started moving off Claude within days, for reasons unrelated to the model, and every provider now publishes retirement dates for the models you depend on. Why the switching cost sits in prompts, evals, traces and data access, and the supplier-risk checklist a CFO and COO can run before the next renewal. Hendrik De Winne 26 October 2026 9 min read In the last week of February 2026, an executive at a US defense technology company told employees to start switching Claude out for other models, including some open-source ones. The executive expected the work to take “a week or two,” according to [CNBC](https://www.cnbc.com/2026/03/04/pentagon-blacklist-anthropic-defense-tech-claude.html). Nothing was wrong with the model. On Friday, February 27, President Trump ordered every federal agency to “immediately cease” using Anthropic’s technology, and Defense Secretary Pete Hegseth said he was directing the Pentagon to designate the company a supply chain risk. Alexander Harstrick of J2 Ventures said ten of his portfolio companies were replacing Claude in their defense work, and that the move “in no way reflected a perceived shortcoming of Claude.” The first thing a CFO should take from that week is that a company can lose its AI supplier through no fault of its own and no fault of the supplier’s. The second is less obvious. The real switching cost sat in prompts tuned to one model, in workflows built on one vendor’s tools, and in the absence of any test that could show a replacement was good enough. Those costs can be cut in advance, cheaply, and almost nobody does it: in a mid-2025 survey by [Menlo Ventures](https://menlovc.com/perspective/2025-mid-year-llm-market-update/), an Anthropic investor, only **11%** of teams building with AI had changed model providers in the past year. So the rule I would use is plain: treat the model provider like a single-source supplier of a critical part. Keep a second source qualified, own the specifications, and write the exit into the contract before you need it. - **88%** Share of enterprise LLM API usage held by Anthropic, OpenAI and Google, per [Menlo Ventures](https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/), December 2025. - **11%** Teams that changed model providers in the past year, mid-2025, per [Menlo Ventures](https://menlovc.com/perspective/2025-mid-year-llm-market-update/). - **60days** Minimum notice [Anthropic](https://docs.claude.com/en/docs/about-claude/model-deprecations) gives before retiring a public model. - **<5days** Notice Windsurf says it got before losing most direct Claude capacity in June 2025, per [TechCrunch](https://techcrunch.com/2025/06/03/windsurf-says-anthropic-is-limiting-its-direct-access-to-claude-ai-models/). ## A supplier can disappear for reasons that have nothing to do with you In a [statement on February 26](https://www.anthropic.com/news/statement-department-of-war), Anthropic CEO Dario Amodei said the Department of War would only contract with AI companies that accept “any lawful use,” while Anthropic held to two exceptions: mass domestic surveillance of Americans and fully autonomous weapons. Trump’s post set a six-month phase-out for agencies such as the Defense Department. On March 4 the Pentagon sent Anthropic a letter [confirming the designation](https://www.anthropic.com/news/where-stand-department-war). By then, CNBC reported, officials at Treasury, State and Health and Human Services had also told employees to move off Claude. The facts deserve to be stated fairly. Anthropic [said](https://www.anthropic.com/news/statement-comments-secretary-war) its commercial customers kept full access, and Microsoft, Google and Amazon all [kept offering Claude](https://www.cnbc.com/2026/03/06/google-says-anthropic-remains-available-outside-of-defense-projects.html) outside defense work. On March 26, Judge Rita Lin of the Northern District of California issued a [preliminary injunction](https://www.cnbc.com/2026/03/26/anthropic-pentagon-dod-claude-court-ruling.html) against the president’s directive and the designation, calling the government’s conduct “classic illegal First Amendment retaliation.” On August 27 she [ruled the designation unlawful](https://www.cnbc.com/2026/08/28/judge-blocks-pentagon-blacklist--anthropic-.html). A parallel case in the D.C. Circuit was still open at that point. None of that helped a contractor in the first week of March. It had to assume the ban would hold and move on a timeline set by a social media post. The lesson applies to every provider. Any of them can be cut off by a government, an acquisition, a capacity squeeze or a pricing decision, and the customer has no say in which one arrives first. ## Every model you test now comes with a retirement date Politics is the rare case. Retirement is the routine one, and the providers publish it. On Anthropic’s [deprecations page](https://docs.claude.com/en/docs/about-claude/model-deprecations), Claude Sonnet 4 and Opus 4 were announced for retirement on April 14, 2026 and switched off on June 15. Opus 4.1 followed on August 5. After that date, the page says, “Requests to models past the retirement date will fail.” OpenAI’s [deprecations page](https://platform.openai.com/docs/deprecations) promises at least six months’ notice for generally available models and as little as two weeks for preview models. On October 14, 2026 it told developers that several GPT-4o and GPT-4.1 mini snapshots leave the API on January 14, 2027. It also retired a whole interface: the Assistants API, OpenAI’s earlier toolkit for building assistants, was removed on August 26, 2026, a year after notice. [Balaji Srinivasan](https://x.com/balajis/status/1931671546113994769) put the economics in one line last year: “AI doesn’t take your job. AI takes the job of the previous AI.” For a buyer, that means the model is the least durable part of any AI investment. A workflow that works today will run on a different model within a year or so, whether you choose the date or the provider does. ## The data your agents read is rented too Most of the value in an agent comes from what it can reach: the CRM, the ticket queue, the ERP, the chat history. Those systems belong to other vendors, and their terms can change. On May 29, 2025, [Slack](https://docs.slack.dev/changelog/2025/05/29/rate-limit-changes-for-non-marketplace-apps) cut the rate limit on the two API methods that read message history to one request per minute, returning at most 15 messages, for new installs of commercially distributed apps outside its Marketplace. Slack said the change was meant to stop “bulk data exfiltration by unvetted applications.” Apps a company builds for itself were exempt. Its updated [API terms](https://slack.com/terms-of-service/api) also bar third-party apps from using Slack data to “train a large language model” or bulk exporting it without a separate agreement. Slack had its reasons. The point for a COO is narrower: if you bought an AI search or assistant tool that reads your Slack, its ability to do that now depends on its standing with Salesforce, which owns Slack. The same applies to every AI tool sitting on top of a system of record you rent. Model access can move just as fast. In June 2025, Windsurf, a coding tool then reportedly being acquired by OpenAI, said Anthropic had cut off “nearly all” of its first-party capacity to Claude 3.x models “with less than five days of notice.” Windsurf had some backup capacity with other inference providers, “but not enough,” and warned its own users of availability problems. That case involved a competitor, but the mechanism carries over: one decision upstream, and a product downstream degrades the same week. ## The switching cost sits in four things you can own When a team says changing models would take months, the time goes into four things, and a company can own all of them. **Prompts and instructions.** The standing instructions an agent follows are business documents: how to code an invoice, which customers get a callback, what a refund requires. They belong in your own repository, with an owner and a review date, in plain text that any model can read. Anthropic published its [Agent Skills format](https://claude.com/resources/articles/organization-skills-and-directory) as an open standard in December 2025, saying “the same skill should work whether you’re using Claude or other AI platforms.” **Evals.** An eval set is a collection of real cases from your own work with the answer you expect: a hundred invoices with the right coding, fifty support tickets with the right next step. It is the only thing that lets you say a second model is good enough, and it turns a migration into a test run. I set out how corrections become those tests in [the note on fixing a mistake once](https://vandam.ai/notes/fix-a-mistake-once). **Traces.** The logs of what each agent read, decided and did are your audit trail and your best source of new test cases. Keep them in your own storage, in a format you can query without the vendor’s console. **Integration code.** Agent frameworks tied to one lab are the costliest thing to leave. A thin layer of your own between the workflow and the model, so the model name is one setting, costs little to build at the start and a lot to retrofit. ## Your AI services vendor should hand back what it learned The same four things matter when someone else builds your agents. A lab’s forward-deployed team, a consultancy or a small firm like ours spends months learning how your work runs, and that learning ends up in prompts, skill files, evals and traces. One builder who posts as [@dom__inic](https://x.com/dom__inic/status/2066260212772679864) argued in June that “organizations are effectively outsourcing key learning curves and domain knowledge to the AI labs.” That is his reading of the labs’ services businesses. The labs’ own terms point the other way on training: Anthropic says that [by default](https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training) it does not train on inputs or outputs from its commercial products, though feedback you send it can be used. Training is only half the concern. The other half is whether you could keep running the work if the vendor left. If the prompts live in the vendor’s workspace, the evals in its notebooks and the traces in its account, you will rent your own know-how back at renewal. Whichever kind of [provider you pick](https://vandam.ai/notes/who-does-the-work), the contract should say that every prompt, skill file, eval set, trace and piece of integration code is yours, delivered in open formats every quarter and in full at exit. That is the standard we hold our own work to, and a provider that will not sign it has told you something useful before the renewal. ## A supplier-risk checklist for the CFO and COO This is the checklist we use. The COO owns the first five items, the CFO the last three. 1. **List every workflow that calls a model,** with the provider, the exact model version and the business process it supports. Mark the ones where an outage would stop revenue, payroll, the close or a customer commitment. 2. **Check each model’s retirement date** on the provider’s own deprecations page and put it in the same calendar as contract renewals. Anything on a preview model in production gets moved or signed off by name. 3. **Qualify a second provider** for each critical workflow. Run the same eval set on both, record the scores, and keep the second one live at small volume so the keys, billing and access already work. 4. **Run a 72-hour switch drill** once a year on one critical workflow: route it to the second provider, rerun the evals, and write down what broke and how long the fix took. 5. **Map the data dependencies.** For each agent, list the systems it reads and writes, whose API that is, and what happens to the workflow if the rate limit drops to one call a minute. 6. **Put data access into renewals.** For every system of record, get written terms on API access for your own agents and for the third-party AI tools you approve, with notice periods for rate limit and terms changes, the way you already negotiate uptime. 7. **Put the handback into services contracts,** as above, and check the model provider’s training defaults and feedback exception. 8. **Price the switch.** Put a number on what one critical workflow would cost to move in a week: engineering hours, review hours, any rate difference and a month of lower quality. Carry it in the risk register next to other single-source suppliers. A second provider costs money and attention, so I would keep one warm only for workflows on the critical list. ## What to do before the next renewal This month, the COO should pick the one AI workflow the company would miss most, find its model version on the provider’s deprecations page, and ask the team how they would know a replacement model is good enough. If the answer is a demo, have them build a fifty-case eval set from last quarter’s real work before doing anything else. Before the next renewal of any system your agents read, the CFO should add one page to the negotiation covering API access, rate limits and notice periods, and before signing any AI services contract, ask for the handback clause in writing. Then schedule the first switch drill for a quiet week, while it is still your choice when it happens. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Your outside firms got faster with AI and you are still paying for the old hours. · VANDAM > Your law firm, auditor and agencies got faster with AI while invoices stayed the same. A CFO playbook to reprice per deliverable and renegotiate software. URL: https://vandam.ai/notes/reprice-your-outside-firms Contact: hendrik@vibelab.be # Your outside firms got faster with AI and you are still paying for the old hours. Law firms, auditors, agencies and outsourcers are putting AI into the work they bill you for, and most contracts still price the old hours. KPMG showed how to ask for the saving back, and I set out the questions a CFO should send each supplier before the next renewal. Hendrik De Winne 26 October 2026 9 min read In 2025, KPMG International went back to its own auditor, Grant Thornton UK, and asked for a cut. The argument, as the [Financial Times](https://www.ft.com/content/c891c47c-b21f-4e0f-84b3-b80c794eff3d) reported in February, was that Grant Thornton was rolling out AI across its audit work and should share the savings. KPMG reportedly said it might look for another auditor if it did not. According to UK filings cited by [TechSpot](https://www.techspot.com/news/111237-kpmg-asked-own-auditor-discount-citing-ai-efficiencies.html), the fee for auditing KPMG International’s 2025 accounts fell from **$416,000** to **$357,000**, about **14%**. KPMG sells audits for a living. It knows better than almost any buyer what AI does to the hours inside one, and it used that knowledge as a client. A mid-sized company typically pays a law firm, an audit firm, a tax adviser, agencies and an outsourcer, and all of them are putting AI into the work they bill you for. Very few of those contracts have changed. The part a CFO may not see from the invoice is who decides where that saving goes. Under hourly billing the supplier decides, quietly, matter by matter. A firm that does ten hours of work in five can bill five hours at a higher rate, bill eight, or move you to a “value-based” fee set by what the work is worth to you. Each of those keeps most of the saving at the supplier. You only get a share if you ask for one, and the best time to ask is the next renewal. ## Your suppliers’ costs fell before your invoices did The legal market shows the gap most clearly. The [2026 Report on the State of the US Legal Market](https://www.thomsonreuters.com/en/institute/reports/state-of-the-us-legal-market-2026), from the Thomson Reuters Institute and Georgetown Law, found that the average US law firm grew profit **13%** in 2025 while worked rates rose a record **7.3%**. It also notes that “90% of legal dollars still flow through hourly billing arrangements that may no longer reflect the value delivered,” despite heavy AI investment. The clients have noticed. In October 2025 the [Association of Corporate Counsel and Everlaw](https://www.acc.com/about/newsroom/news/acc-genai-report-corporate-law-departments-ai-use-everlaw) surveyed 657 in-house legal professionals in 30 countries. Nearly 60% reported “no noticeable savings yet” from their outside counsel using generative AI. Asked why, 58% said law firms have not adjusted pricing to reflect the efficiency. And 59% did not know whether their law firms used generative AI on their matters at all. - **90%** Share of US legal dollars still billed by the hour, per the [Thomson Reuters Institute](https://www.thomsonreuters.com/en/institute/reports/state-of-the-us-legal-market-2026), January 2026. - **58%** In-house lawyers who say firms have not adjusted pricing for AI, per [ACC and Everlaw](https://www.acc.com/about/newsroom/news/acc-genai-report-corporate-law-departments-ai-use-everlaw), October 2025. - **59%** In-house teams that do not know whether their law firms use AI on their matters, same survey. - **14%** Cut in KPMG International’s audit fee after it asked for AI savings, per the [FT](https://www.ft.com/content/c891c47c-b21f-4e0f-84b3-b80c794eff3d). The size of the possible saving is still an estimate. Clio, which sells practice software to law firms, writes in its [Legal Trends research](https://www.clio.com/blog/death-of-the-billable-hour-legal-ai/) that as much as **74%** of billable work could be automated with AI. That is a vendor’s estimate of potential. Thomson Reuters’ [Future of Professionals](https://www.thomsonreuters.com/en/press-releases/2025/june/the-ai-adoption-reality-check-firms-with-ai-strategies-are-twice-as-likely-to-see-ai-driven-revenue-growth-those-without-risk-falling-behind) survey of 2,275 professionals in 2025 is more modest: respondents predicted that professionals using AI will save **5 hours** a week within a year. That is more than 10% of a working week, on work you pay for by the hour. ## The firms are already asking who keeps the saving The accounting firms say this openly. In March, [Bloomberg Tax](https://news.bloombergtax.com/financial-accounting/ai-efficiency-gains-push-accounting-firms-to-reimagine-pricing) asked partners at KPMG, PwC and RSM how AI changes what they charge. RSM, which focuses on mid-sized companies, gave a concrete example: a tax advisory report that took a team 20 hours could take 10 with generative AI and a human check of the output. Sergio de la Fe, the RSM partner who leads its enterprise digital work, put the dilemma in one question. Is that all margin for me or is that all price reduction for our client? (Sergio de la Fe, partner at RSM US, to [Bloomberg Tax](https://news.bloombergtax.com/financial-accounting/ai-efficiency-gains-push-accounting-firms-to-reimagine-pricing), March 2026) His answer was that RSM can cut prices in some areas and not in others. PwC’s US tax leader, Krishnan Chandrasekhar, told the same reporter that “time’s becoming less and less of relevance,” and described pricing some work on the value the client ascribes to it. KPMG’s Rob Fisher called value-based pricing “incredibly important” and pointed out that AI tooling is a cost that is hard to attribute to individual projects. RSM announced a **$1 billion** AI investment over three years, so the firms have real costs of their own to recover. Agencies are further along, and some founders say so in public. [Jake Ward](https://x.com/jakezward/status/2044762030758928467), who owns an SEO agency, described the shift in April: the old model was an “Army of humans” at “30% margins,” and the new one is “One person does the work of seven” at “75%+ margins.” Those are his own figures, posted to promote the model, so I read them as a claim. They still tell a buyer something useful. The agency on the other side of that post keeps selling “the outcome on a monthly retainer,” and the retainer was priced when the work took seven people. ## Value pricing set by the supplier keeps the saving there Every one of these firms is moving away from the hour, which sounds like what the client wants. The detail that matters is the anchor. A value-based fee is set by what the work is worth to you, and that number did not fall when the hours did. If the firm proposes the new price, it will start from the old invoice. The buyer’s version is a price per deliverable anchored to what the deliverable used to cost, with an agreed share of the saving coming back each year. A fixed fee for each NDA, each quarterly tax provision, each monthly paid-search report or each processed invoice. The supplier keeps the upside from getting faster, which gives it a reason to keep using AI. You get a lower unit price and a clear definition of done. The rule we use: take the average cost of one deliverable over the last twelve months as the ceiling, ask for 20 to 40% of the estimated saving back in the first year, and write in a review of the price every twelve months. The exact share is negotiable, as long as the contract states one. ## Software renewals are the same negotiation The same logic applies to the software line, from the other direction. Zylo, which sells SaaS management software, says in its [2026 SaaS Management Index](https://zylo.com/blog/saas-usage-numbers) that the average organization uses **54%** of the licenses it provisions. Its [report page](https://zylo.com/reports/2026-smi/) adds that 78% of IT leaders reported unexpected charges tied to AI features or consumption-based pricing. Those are figures from a vendor that profits when you worry about waste, so treat them as indicative. What has changed is the buyer’s alternative. When [Greg Isenberg](https://x.com/gregisenberg/status/2105397216374403561) read Google’s September announcement that its agents were rewriting old C and C++ code, he wrote: “There are PLENTY of companies stay with outdated software vendors only because switching is painful, and once rewriting is cheap, that lock in starts to break!” A tool that 20 people use for a tenth of its features can now be rebuilt internally or swapped for something cheaper, and the vendor’s account manager knows it. I made the general build-or-buy case in [the note on Coase and the firm](https://vandam.ai/notes/coase-and-the-firm). At renewal, it becomes a specific number: what replacing this tool would cost you. Keep the systems of record out of this. The ERP, the general ledger, the CRM and the HR system are what your agents will read from and write to, and they get more valuable as agents arrive. The candidates are the seat-priced tools on top. ## The questions to ask each type of supplier These are the questions I would send before each renewal, in writing, with a request for written answers. **Law firms.** Which tasks on our matters do you now do with AI, and with which tools? How has the time per task changed for contract review, first drafts, research and discovery since 2024? Will you quote a fixed fee per standard contract, per NDA and per routine filing, based on last year’s average? What is your policy on billing hours an AI tool did not need? Who reviews AI output before it reaches us? **Audit and accounting firms.** Which audit procedures now test the full population with automation, and which still sample? How many hours did last year’s engagement take, by phase, and what do you expect this year? What share of the efficiency is reflected in the fee? For tax and advisory work, will you price per return, per provision and per memo? Auditors have to stay independent and the quality bar is set by regulators, so frame this as a fee conversation and leave the audit scope to them. **Agencies.** Which deliverables are now produced mainly by AI with a person checking them? How many people work on our account today compared with when the retainer was set? Will you price per deliverable or per result we can count, such as qualified leads, published pages or campaigns live? Who owns the prompts, templates and tools built on our account if we leave? **Outsourcers and managed services.** For BPO, IT support or managed finance: what share of tickets, invoices or transactions is now handled without a person? Will you move from per-FTE pricing to a price per resolved ticket or processed document, with a quality floor? What service levels change, and what happens to the price when automation rises next year? **Software vendors.** How many of our seats were active in the last 90 days? Can we drop to the active count at renewal? What AI features are now bundled into the price, and can we decline them? What would it cost to leave, including data export? ## Where this can go wrong Pushing price without watching quality is the obvious risk. In October 2025, [the Associated Press](https://www.news4jax.com/business/2025/10/07/deloitte-to-partially-refund-australian-government-for-report-with-apparent-ai-generated-errors/) reported that Deloitte would partially refund the Australian government for a A$440,000 report with incorrect references and a fabricated quote. The revised version disclosed that Azure OpenAI was used in writing it. A price per deliverable needs a definition of an acceptable one, a right to reject, and a clause requiring the supplier to disclose AI use on your work. ACC’s 59% figure suggests most contracts have none of these. The second risk is squeezing a supplier you depend on. The firms are spending heavily on AI, and Grant Thornton told the FT, as TechSpot quoted it, that its fees “reflect both the cost of our people and the cost of the technology that supports them.” That is fair. A share of the saving is a different request from all of it, and a supplier that keeps some of the gain will keep investing in getting faster. The third is doing this before you know your own numbers. If you cannot say what one contract review or one monthly report cost you last year, the supplier sets the anchor. The unit-cost discipline from [the note on pricing AI per finished task](https://vandam.ai/notes/cost-of-the-i) applies to outside firms as much as to your own AI budget. ## What a CFO can do this quarter Pull the last twelve months of spend with outside firms and software vendors, and rank the suppliers by total. Put every contract over $50,000 on one renewal calendar with its notice date. For the top ten, count the deliverables you received and divide to get a cost per deliverable. Send each of them the questions above, adapted to the supplier type, at least 90 days before renewal. Then choose three contracts, one firm, one agency or outsourcer and one software tool, and convert them to a price per deliverable with a stated share of the saving and an annual review. The quarter after that, you will know which suppliers are sharing the gain and which are keeping it. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # How to tell your people what AI will change about their jobs. · VANDAM > How CEOs and CHROs should tell employees what AI changes: what Amazon and Duolingo show, the trust data, a six-step sequence and what never to promise. URL: https://vandam.ai/notes/telling-your-people Contact: hendrik@vibelab.be # How to tell your people what AI will change about their jobs. Most employees can see AI arriving, and only a quarter say their employer has explained a plan. What silence and shifting explanations cost, a communication sequence for the CEO and CHRO, and what never to promise. Hendrik De Winne 25 October 2026 8 min read On June 17, 2025, Andy Jassy sent Amazon’s employees a memo about generative AI. It was candid. “We will need fewer people doing some of the jobs that are being done today,” he wrote, and “more people doing other types of jobs.” Over the next few years, he expected AI to “reduce our total corporate workforce.” He even admitted what he did not know: “It’s hard to know exactly where this nets out over time.” You can read it on [About Amazon](https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-on-generative-ai). Four months later, Amazon cut about **14,000** corporate roles. The message from Beth Galetti, who runs people experience and technology, explained the cut through layers and bureaucracy, and added that “this generation of AI is the most transformative technology we’ve seen since the Internet,” according to [Amazon](https://www.aboutamazon.com/news/company-news/amazon-workforce-reduction). Two days later, on the earnings call, Jassy told analysts the cut was “not really financially driven, and it’s not even really AI-driven, not right now at least. Really, it’s culture,” as [GeekWire](https://www.geekwire.com/2025/its-culture-amazon-ceo-says-massive-corporate-layoffs-were-about-agility-not-ai-or-cost-cutting/) reported. In January 2026 came another **16,000** roles, with a [message](https://www.aboutamazon.com/news/company-news/amazon-layoffs-corporate-jan-2026) that did not mention AI at all and answered the obvious fear directly: some employees might ask whether broad reductions every few months were the new rhythm. “That’s not our plan.” Each of those messages can be defended on its own. Read in sequence by someone who works there, they say four different things about why jobs are going. I start with Amazon because it is a company with every resource for internal communication, and it still produced the pattern I see in mid-sized firms: an honest first message, followed by explanations that drift. Employees keep the first one and judge everything after it against that. - **52%** US workers who feel worried about future AI use in the workplace, per [Pew](https://www.pewresearch.org/social-trends/2025/02/25/workers-views-of-ai-use-in-the-workplace/), surveyed October 2024. - **25%** US employees who say their organization has communicated a clear plan for AI, per [Gallup](https://www.gallup.com/699797/indicator-artificial-intelligence.aspx), May 2026. - **66%** Employees who trust their own CEO, down from 70% and a record low, per [Edelman](https://www.edelman.com/newsroom/richard-edelmans-6am-blog/trust-at-work-2026), October 2026. - **18%** Employees with a negative transformation experience who would embrace more AI at work, against 40% after a positive one, per [Edelman](https://www.edelman.com/newsroom/richard-edelmans-6am-blog/trust-at-work-2026). ## Most employees have heard nothing they can use The worry is broad and it predates most of the memos. When [Pew Research Center](https://www.pewresearch.org/social-trends/2025/02/25/workers-views-of-ai-use-in-the-workplace/) surveyed 5,273 employed Americans in October 2024, **52%** said they felt worried about how AI will be used at work and **36%** felt hopeful. **32%** expected fewer job opportunities for themselves in the long run, and only 6% expected more. Lower-income workers were the most pessimistic. Use was still thin: by September 2025, about **21%** of workers said at least some of their work was done with AI, up from 16% a year earlier, per a [Pew follow-up](https://www.pewresearch.org/short-reads/2025/10/06/about-1-in-5-us-workers-now-use-ai-in-their-job-up-since-last-year/). So most people form their view of AI at work from headlines about other companies. Their own employer rarely fills the gap. In [Gallup’s](https://www.gallup.com/699797/indicator-artificial-intelligence.aspx) May 2026 reading, **47%** of US employees say their organization has integrated AI, and only **25%** say it has communicated a clear plan. That leaves a large group of people who can see the tools arriving and have been told nothing about what they mean for their job. Edelman’s 2026 [Trust at Work](https://www.edelman.com/newsroom/richard-edelmans-6am-blog/trust-at-work-2026) study, published this month from a survey of more than 10,000 employees in ten markets, shows what that silence lands on. AI implementation is now the most common transformation employees report, at **53%**. Three quarters of workers fear losing their job to tariffs, automation, offshoring or a lack of training. Trust in “my CEO” fell from 70% to **66%**, and trust in direct managers from 79% to 72%, each a record low in Edelman’s series. Associates were roughly three times as likely as executives to call their experience of the change less than positive, 49% against 17%. The people deciding and the people affected are living through different events. ## Uncertainty does more damage than bad news The research on this is older than generative AI, and it is consistent. In a field experiment published in 1991, [David Schweiger and Angelo DeNisi](https://doi.org/10.5465/256304) followed two plants of a company that had just announced a merger. One plant received what they called a realistic merger preview, a program of frank and regular communication about what would happen. The other received limited information. Across four rounds of measurement, the preview reduced the damaging effects of the merger, and in some cases attitudes returned to where they had been before the announcement. Both plants went through the same merger, and the plant that heard about it early and often from the company came through it in better shape. The cost of getting it wrong shows up in who leaves. [Charlie Trevor and Anthony Nyberg](https://doi.org/10.5465/amj.2008.31767250) studied organization-level data across industries in 2008 and found that downsizing predicts higher voluntary turnover among the people who stay, through lower commitment. HR practices that convey procedural fairness or tie people to the organization softened that effect. My reading, which goes beyond their data: the people who leave voluntarily are the ones with options, and in an AI transition those are usually the people already doing the most with the tools. Edelman’s new numbers say the same about AI specifically. Employees with a negative experience of transformation were less than half as likely to embrace more AI at work, **18%** against **40%**. Commitment to a long-term career at the company was 51% after a negative experience and **83%** after a positive one. A badly communicated rollout makes the next rollout harder, because the people you need to adopt it have stopped believing you. ## Broken promises do the most damage Most damage comes from a promise or an explanation that later turns out to be incomplete. Duolingo is the clearest case. Luis von Ahn’s April 2025 “AI-first” memo said the company would gradually stop using contractors for work AI could handle, as [Fortune](https://fortune.com/2025/06/09/duolingo-ceo-surprised-backlash-ai-first-company-announcement/) reported. Readers outside the company heard layoffs. Months later he was still correcting the record. > We’ve never laid off any full-time employees. We don’t plan to. > > Luis von Ahn, CEO, Duolingo, August 2025, [Source](https://fortune.com/2025/08/18/duolingo-ceo-admits-controversial-ai-memo-did-not-give-enough-context-insists-company-never-laid-off-full-time-employees) Look at the second sentence. It is the kind of line every CEO is tempted to say, and the one that costs the most if circumstances change. Amazon’s “That’s not our plan” has the same shape. Both are probably sincere. Neither is a commitment a CEO fully controls, because markets, models and boards move. The memos from Shopify and Duolingo about AI use, which I wrote about in [an earlier note](https://vandam.ai/notes/tinkerers), ran into a related problem: when the message is mainly an instruction to use AI, people start asking what the instruction is really for. The CEOs who did this better, in my reading, did two things. They said plainly which part was uncertain, as Jassy did in June. And they kept the explanation stable afterwards. Amazon’s June memo would have aged well if October and January had repeated its logic. The memos went wrong in the follow-up. ## A communication sequence that holds up This is the order we use with clients. It assumes a CEO and a CHRO doing it together, with the CFO in the room for the first step. 1. **Before saying anything, write three lists.** What we know (which tasks agents already do, in which teams), what we have decided (budget, tools, the first roles to be redesigned), and what we have not decided (headcount, timing, which roles shrink). The third list is the one most leadership teams skip, and it is the one employees care about. If headcount reduction is on the table, it goes on that list honestly. 2. **Week one: the CEO tells everyone at once, in person or live.** Give the business reason in terms of competitiveness and customers. Edelman reports that employees see change as justified by outcomes like competitiveness and survival, and much less by shareholder value. Read out all three lists. End with the date of the next update. 3. **Within 48 hours: every manager holds a team conversation.** Give them a one-page brief and the three lists the day before the CEO speaks. [Gallup](https://www.gallup.com/workplace/712433/employee-engagement-remains-flat-adoption-accelerates.aspx) finds engagement of 48% among employees whose manager actively supports AI use against 30% among those whose manager does not, and **53%** when frequent use, a clear plan and active manager support come together. The manager is the channel people believe. 4. **First month: role-by-role conversations.** For each role, write down which tasks move to agents, which new tasks arrive, what training comes with it and who decides. Then talk to the people in that role before the description is final. This is where the redesign work from [the org chart note](https://vandam.ai/notes/org-chart-redrawn) meets real people. 5. **Monthly: update on the date you promised.** Even when nothing changed. Start each update with what you said last time and what actually happened. Move items from “not decided” to “decided” in public. A steady cadence does more for trust than any single speech. 6. **If roles go, say it once, with the real reason, and keep it.** State the criteria, the notice, the support and the internal options. Amazon’s 90 days to look for an internal role is a mechanism worth copying. Then give the same explanation to analysts, press and staff, because they compare notes. What never to promise: that no jobs will be lost, that a given role will stay as it is, a date you do not control, or a reason you might contradict next quarter. What you can promise and keep: how decisions will be made, how much notice people get, that the people affected will be consulted first, that they get training before the switch, and first look at new internal roles. ## What to do this month Sit down with your CHRO and write the three lists for your own company, on one page. If the “not decided” list is empty, you have probably decided more than you have told anyone, or less than you think. Then make a second list of the twenty people you can least afford to lose, especially the ones already doing the most with AI. Have each of their managers talk with them within two weeks about what their job becomes, before the all-hands. They are the first to hear the rumor and the first to get a call from a recruiter. Finally, put the first all-hands and the next three monthly update dates in the calendar now, and send the dates to the whole company. The calendar invite is the first promise, and it is one you can keep. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # AI transformation has ten steps and only one of them is AI. · VANDAM > The ten steps of making a process AI-native, walked through one quote process: why mapping, data and adoption take the time and the AI step is short. URL: https://vandam.ai/notes/ten-steps-one-is-ai Contact: hendrik@vibelab.be # AI transformation has ten steps and only one of them is AI. Alex Lieberman of Tenex listed the ten steps it takes to make a process AI-native. I walked them through one quote process at a parts distributor to show where the weeks go, and why the map pays off before any model does. Hendrik De Winne 25 October 2026 7 min read In May, [Alex Lieberman](https://x.com/businessbarista/status/2054684459279012069), co-founder of Morning Brew and of [Tenex](https://www.tenex.co), an AI transformation firm, put my whole job in one sentence: “AI transformation is just a cooler way of saying data engineering, process mapping, change management and intelligent automation.” A month later [he spelled it out](https://x.com/businessbarista/status/2064089261121626531). “There are 10 steps in the sequence of making an internal process or external product AI-native. Only 1 step is AI, and ironically, the other 9 steps are the far harder part.” His list is the best checklist I have seen for this work, so I want to walk it through one process from start to finish and show where the calendar actually goes. The process is an illustration, built from how this kind of work usually runs, and no client is behind it. Take an industrial parts distributor with a few hundred people. Customers send quote requests by email, as typed lists, PDFs or photos of a worn part. An inside-sales rep reads each one, translates the customer’s own part names into catalog numbers, checks stock in the ERP, looks up the customer’s negotiated prices in a spreadsheet the sales manager keeps, asks a product specialist when a part is discontinued, types the quote into the ERP, sends a PDF and follows up a few days later. Everyone agrees quotes are too slow. Nobody can say exactly why. ## The first three steps take most of the time 1. **Identify the problem.** Lieberman’s instruction is to “find the manual process worth automating” and to turn on your “suck meter.” For the distributor that part is easy, because sales already complains about it. The harder part is writing the goal down. Tenex teaches a template for this in [its workshop for CFOs and CHROs](https://x.com/businessbarista/status/2101659025788031152): “Improve [measure] from [baseline] to [target] by [date].” Here that becomes: improve median quote turnaround from whatever it is today to the same day, by the end of next quarter. Nobody knows the baseline yet, which tells you what step two is for. 2. **Understand the workflow.** “Map how people actually work today,” on paper or a whiteboard, from beginning to end. Lieberman calls it the least sexy part and says it is “generally where the people driving transformation (FDE, GTM engineer, etc) should spend the majority of their time.” At the distributor that means sitting next to three reps for a few days and timing every hop. The map usually surprises people. In this example it shows that the slow step is the price spreadsheet: only the sales manager edits it, and reps wait for his answer before they can send anything. It also shows that a large share of requests repeat an order the customer placed before. 3. **Collect the data.** “Gather sample inputs, documents, edge cases.” Here that means a few hundred past requests together with the quotes that answered them, the photos nobody could read, the discontinued parts and every customer who uses their own codes. The most valuable thing this step produces is a table that maps each customer’s part names to catalog numbers. Right now it lives in the heads of the two longest-serving reps. Two fixes fall out of the map before anyone writes a prompt. Agreed prices move from the spreadsheet into the ERP, where any rep can read them, and a repeat request can be answered by copying the last quote. Lieberman notes this in his first step: “funny enough, your company becomes more efficient just by mapping out your processes even if you don’t introduce AI.” In [an April training](https://x.com/businessbarista/status/2047488697885413391) for executives from a large financial institution, his team had every exec write out a key process and mark each step E (eliminate), A (automate) or D (delegate). He reports it “revealed opportunities for AI, but also existing inefficiencies that have gone undetected.” This is old knowledge. In 1990, [Michael Hammer](https://hbr.org/1990/07/reengineering-work-dont-automate-obliterate), then president of the consulting firm Hammer and Company, described how Ford looked at its North American accounts payable department of more than 500 people. Management hoped new computers would cut headcount by some 20%. Then it saw that Mazda ran accounts payable with a total of 5 people. When Ford mapped the old process, it found the department “spent most of its time on mismatches” between purchase orders, receiving documents and invoices. Ford stopped asking for invoices altogether. - **143** Data items accounts payable had to match before paying a vendor, before and after Ford’s redesign, per [HBR](https://hbr.org/1990/07/reengineering-work-dont-automate-obliterate). - **75%** Headcount reduction where the new process ran, against the 20% the original plan aimed for. It is time to stop paving the cow paths. (Michael Hammer, [Harvard Business Review](https://hbr.org/1990/07/reengineering-work-dont-automate-obliterate), July 1990) ## The AI step is the short one 1. **Build the prototype [The AI Part].** Lieberman’s brief is to “just worry about proving there’s a there there.” For the distributor, the prototype is an agent that reads a request in any format, matches each line to a catalog number using the table from step three, pulls stock and price, and drafts a quote for the rep to approve. With the map and the data in hand, this is the step measured in days. Without them, it is where projects go to die, because nobody can say what a correct quote looks like. > Before you reimagine a process you have to become an expert in that process. > > Alex Lieberman, Co-founder, Tenex, June 2026, [Source](https://x.com/businessbarista/status/2064089261121626531) ## From one user to a whole sales team 1. **Test & iterate.** “Validate with real users and edge cases.” The past requests from step three become the test set, because the right answer is already known. A rep runs the agent next to their normal work for a few weeks and corrects every draft. One question from [the list his team collects from enterprise executives](https://x.com/businessbarista/status/2077143133049995619) belongs here: “When an agent does eight hours of work, how does a human check it in eight minutes?” For quotes, the answer is a draft that shows which lines it was unsure of, so the rep checks three lines and skips the forty it matched cleanly. 2. **Integrate with systems.** “Point-in-time data is good for testing the workflow, but live data is necessary before going into production.” Now the agent needs read access to the ERP, the shared mailbox and the CRM, and someone in IT has to approve each of them. [Aaron Levie](https://x.com/levie/status/2067455756795039957), CEO of [Box](https://www.box.com), named this work in June as a big reason the applied AI layer works at scale: “Data has to be cleaned up and moved to modern systems, processes have to be re-engineered and documented, workflows have to be evaled, SLAs have to get achieved, and so on.” 3. **Roll out & train.** The next step is “hand-holding your peers/users through the onboarding process.” At the distributor that is twelve reps, each with their own habits and a few customers they guard closely. ## The last three steps are where companies stall Lieberman also posted [notes from a call](https://x.com/businessbarista/status/2056879209583349801) with the Chief HR Officer of a $500m consumer business, and they read like a description of a company that stopped after step four. Everyone has ChatGPT access and basic prompting training. A smaller group has Claude with deeper permissions. The CHRO’s biggest worry is “a widening gap between power users and the rest of the org,” and the most advanced users are “building agents for personal workflows, not yet for scalable company processes.” 1. **Drive adoption.** His rule is to introduce a workflow that saves time “and integrates with their already existing behavior so they don’t have to deal with re-education.” For the reps, the drafted quote lands in the same inbox where the request arrived, ready to approve. They never open a new tool. 2. **Empower contribution.** “Treat your new process like an opensource project.” Every time a rep corrects a match, the correction goes into the customer part-name table. The two senior reps who carried that table in their heads now maintain it for everyone, which is a better job than retyping PDFs. 3. **Measure & capture value.** Lieberman is relaxed about ROI while a company is still experimenting. In the scale-up phase he is blunt: “you need to reskill employees attached to this process, undershoot your approved hiring roadmap, or measurably increase ACV/conversion rate/sales cycle speed.” For the distributor this is where the template from step one pays off. The baseline was measured in step two, so the team can show the change in turnaround and decide what the saved rep hours are for: more customers per rep, or fewer new hires next year. Counted honestly, the AI step is a short stretch in the middle. Mapping, data, integration and getting twelve people to change a habit take up the rest, and the first fix, prices in the ERP, ships before any model is involved. That is why I start every engagement with the map and keep the build small until the baseline exists. [The team that does that work](https://vandam.ai/notes/ai-needs-hr-and-finance) needs people who can sit next to a rep for three days and come back knowing where the time goes. If you want to try it on your own company this week, take one process your team complains about, write every step on a single sheet of paper, and put an E, an A or a D next to each one. Some of the Es you can fix on Monday. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # AI spreads through a few people who build for their own work. · VANDAM > AI spreads through a few people who build for their own work. What the Shopify and Duolingo memos show, and what a company can give those people. URL: https://vandam.ai/notes/tinkerers Contact: hendrik@vibelab.be # AI spreads through a few people who build for their own work. Shopify and Duolingo both told every employee to use AI, and Duolingo dropped its usage metric a year later. The people who actually spread AI build tools for their own work, and a company can give them time, status, budget and a way to share what works. Hendrik De Winne 25 October 2026 7 min read In March 2025, [Tobi Lütke](https://x.com/tobi/status/1909231499448401946), the CEO of [Shopify](https://www.shopify.com), sent his staff a memo. In April he heard it was leaking and posted it on X himself. Most of the coverage went to its bold heading, “Reflexive AI usage is now a baseline expectation at Shopify,” and to the rule that teams must show why AI cannot do the work before they ask for more headcount. The sentence I keep going back to sits in the middle. The summer before, Lütke had invited everyone to tinker with AI. Many did, and he describes the people who took it furthest with real admiration. Then he writes: “The call to tinker with it was the right one, but it was too much of a suggestion.” > I’ve seen many of these people approach implausible tasks, ones we wouldn’t even have chosen to tackle before, with reflexive and brilliant usage of AI to get 100X the work done. > > Tobi Lütke, CEO, Shopify, March 2025, [Source](https://x.com/tobi/status/1909231499448401946) Read that way, the memo describes how AI had already spread inside Shopify. A group of people played with it until they were doing work nobody had planned for, and the memo is the company trying to make that everyone’s job. It also builds a way to pass the lessons around. Point four reads “Learning is self directed, but share what you learned,” and names the Slack channels where people already swapped prompts, #revenue-ai-use-cases and #ai-centaurs, with a promise to share “Ws (and Ls!)” and to make time for AI in monthly business reviews. ## Duolingo measured the usage and then let it go Three weeks after Shopify’s memo went public, on April 28, 2025, [Luis von Ahn](https://fortune.com/2025/06/09/duolingo-ceo-surprised-backlash-ai-first-company-announcement/), the CEO of [Duolingo](https://www.duolingo.com), posted his own “AI-first” memo on LinkedIn. According to [Fortune](https://fortune.com/2025/08/18/duolingo-ceo-admits-controversial-ai-memo-did-not-give-enough-context-insists-company-never-laid-off-full-time-employees), it said the company would gradually stop using contractors to “do work that AI can handle,” and that headcount would grow only if “a team cannot automate more of their work.” Employees would also be assessed on their AI use. The reaction came mostly from outside the company. Commenters wrote things like “AI first means people last,” and von Ahn later told the Financial Times, “I did not expect the amount of blowback.” By August he was telling the New York Times, “This was on me. I did not give enough context.” The part that interests me came a year later. In April 2026 von Ahn said on the Silicon Valley Girl podcast that Duolingo had stopped evaluating employees on their AI use. As [Fortune](https://fortune.com/2026/04/13/duolingo-ceo-luis-von-ahn-ai-usage-requirement-employee-performance-evaluations/) reported it, people had started asking whether the company “just wanted them to use AI for AI’s sake.” His new rule: “the most important thing in your performance is that you are doing whatever your job is as well as possible. A lot of times AI can help you with that.” And then: “I’m not going to force you to do that.” Both memos told every employee to use AI. A metric on usage produces usage, and people noticed that the usage was the thing being scored. That lesson cost Duolingo a year. What neither memo can do on its own is turn usage into changed work, and the people who do that turning are a much smaller group. Vas, the CEO of Varick Agents, put that group at **5‑10%** of one large rollout, and I wrote about his barbell in [Why AI licenses rarely make a company faster](https://vandam.ai/notes/adoption-vs-diffusion). This note is about that slice: who they are, and what a company can do so that their work spreads. ## The people who spread it build for their own work first The clearest advice I have read on this came in January from [Nikunj Kothari](https://x.com/nikunj/status/2016532998544560376), a partner at [FPV Ventures](https://fpvventures.com), in a post addressed to CEOs. > Hire tinkerers. Make it high status. Find ones that will explore the edges and (ideally) naturally gifted at teaching people. Empower them to freely roam across the org and fix large problems that can be automated. > > Nikunj Kothari, Partner, FPV Ventures, January 2026, [Source](https://x.com/nikunj/status/2016532998544560376) He also predicts the objection: “All your execs will complain that these tinkerers don’t understand scale or systems (rollout being a favorite word).” His answer is to listen and ignore them, “or even better give them budget to hire a tinkerer to achieve their targets.” [Cody Schneider](https://x.com/codyschneider/status/1993817048896426147), who is building [Graphed](https://www.graphed.com/), describes the habit from the inside: “whatever I’m working on, I’m trying to automate parts of it in the background while I work on it.” After a while, he writes, “you start seeing every repetitive action, every manual process, every annoying workflow as a bug.” His estimate is that fixing “3–5 of these bugs every week” leaves you with your own automations, research agents and monitoring a few months later. He calls the result “the equipped employee,” someone who shows up “with their own internal R&D department in their backpack.” The third example shows what happens when a team gets the same access at once. [The Boring Marketer](https://x.com/boringmarketer/status/1921992953116975536), a marketer who posts under that name, describes an ad agency with one Claude team account connected to “the 3-5 software tools being used on a daily/weekly basis.” “Each person has access. No gatekeepers.” Ad managers write new ad scripts from the agency’s own data without waiting on the Creative Lead, with this result: “Learnings within the day vs a week to get an ad created.” [Aaron Levie](https://x.com/levie/status/2043883641366032638), the CEO of [Box](https://www.box.com), wrote a job description for the formal version of this person in April. He calls it the agent deployer: someone who maps a team’s workflows, finds where agents add the most, and runs them. Two details in it matter here. The person needs “full autonomy to connect up business systems,” and “the company believes it’s safe for them to do so.” And there will be “one or more of these people on every team, so it’s not a centralized role per se.” Put the four next to each other and the pattern is plain. These people start from their own work and their own annoyances. They need permission to connect real systems, which a license alone does not give them. They live inside the teams. And the good ones teach, which is the part Kothari puts in brackets and I would put first. ## Private gains stay private The risk with tinkerers is that their gains stay with them. [Alex Lieberman](https://x.com/businessbarista/status/2056879209583349801), a founder of [Tenex](https://tenex.co) and Morning Brew, shared his notes from a call with the Chief HR Officer of a $500m consumer business. Two lines in them describe most companies I know. “Key challenge is ‘haves vs have-nots’.” And: “Most advanced users building agents for personal workflows, not yet for scalable company processes.” [Seth Rosen](https://x.com/sethrosen/status/2066534763511316540), co-founder of TopCoat and now building Thruwire, gave the pattern a name in June. “If every employee has their own chat, their own memories, and their own files, a few of them get sharper in private. The organization gets nothing.” He calls it “AI hero ball,” and his fix is a contract: everyone keeps their own tools, but every session “leaves the shared structure in better shape than it started.” Reward the people who build the system, and the ones who use it. ([Seth Rosen](https://x.com/sethrosen/status/2066534763511316540), June 2026) ## Give them time, status, access and a way to share A company can design for this with four choices, and I would make them in this order. 1. **Time that is protected.** A tinkerer who builds only in the evenings builds for themselves. Give the people who already experiment a fixed share of their week for it, and put AI on the agenda of the monthly review, the way Shopify’s memo does. 2. **Status that others can see.** Kothari’s “make it high status” is the cheapest item on this list. Let the person who rebuilt the quote process show it at the all-hands, put their name on it, and let them report to someone senior enough to clear the path. 3. **Budget and real access.** Licenses are the easy part. The scarce thing is permission to connect the CRM, the shared drive and the ERP, with someone in IT who agrees it is safe, as in Levie’s description. Without that access the best people stop at copy and paste. 4. **A way to share what works.** A channel where wins and failures get posted, a shared folder of skills and prompts the next person can reuse, and a person who decides which personal tool is worth turning into a company process. That last step is where a tinkerer’s side project becomes a redesigned workflow with an owner. Then score the outcome of the work and leave the usage alone. Duolingo spent a year learning that, and von Ahn’s new rule is the one I would start with. This is also what the [100xworker Program](https://100xworker.com) I run is built around: people who want to become the AI person in their company, and who learn to build for their own work first. If you want a first step this week, find the five people in your company who built something for themselves with AI this year. Ask each of them what they built, who else uses it, and what stopped it from spreading. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # What managers do now that part of every team is agents. · VANDAM > Companies are cutting manager layers as agents take on coordination. What the data says about span of control, and a manager role redesigned for agents. URL: https://vandam.ai/notes/what-managers-do-now Contact: hendrik@vibelab.be # What managers do now that part of every team is agents. Amazon, Microsoft and Meta cut management layers while agents took on coordination work, and the managers who remain carry wider spans and a second team to supervise. What Gallup and Anthropic's data say about that job, and a manager role redesigned around it. Hendrik De Winne 25 October 2026 9 min read In September 2024, Andy Jassy told Amazon’s corporate staff what too many managers had done to the company. Growth had added layers, and the layers had produced “pre-meetings for the pre-meetings for the decision meetings,” a longer line of managers who felt they had to review a topic before it moved, and owners who stopped making recommendations “because the decision will be made elsewhere.” His fix was a number. Every senior leadership organization had to [raise its ratio of individual contributors to managers](https://www.aboutamazon.com/news/company-news/ceo-andy-jassy-latest-update-on-amazon-return-to-office-manager-team-ratio) by at least **15%** by the end of the first quarter of 2025. He also opened a “Bureaucracy Mailbox” for anyone who spotted process worth killing. Thirteen months later, Amazon’s head of people, Beth Galetti, [announced](https://www.aboutamazon.com/news/company-news/amazon-workforce-reduction) a reduction of about **14,000** corporate roles. Her reasoning used the same words: the company needed to be “organized more leanly, with fewer layers and more ownership.” Few CEOs have worked out what the managers who remain are supposed to do all day, once part of each team is software. - **15%** Minimum rise in the ratio of individual contributors to managers [Amazon](https://www.aboutamazon.com/news/company-news/ceo-andy-jassy-latest-update-on-amazon-return-to-office-manager-team-ratio) required by the end of Q1 2025. - **12.1** Average direct reports per US manager in 2025, up from 10.9 a year earlier, per [Gallup](https://www.gallup.com/workplace/700718/span-control-optimal-team-size-managers.aspx). - **22%** Share of managers worldwide who were engaged at work in 2025, down from 27%, per [Gallup](https://www.gallup.com/workplace/349484/state-of-the-global-workplace.aspx). - **20%** Organizations that [Gartner](https://www.gartner.com/en/newsroom/press-releases/2024-10-22-gartner-unveils-top-predictions-for-it-organizations-and-users-in-2025-and-beyond) expects to use AI to cut more than half their middle management roles, through 2026. ## The largest employers cut the same layer Amazon was not alone. At Microsoft, CFO Amy Hood told investors on the [April 2025 earnings call](https://www.microsoft.com/en-us/investor/events/fy-2025/earnings-fy-2025-q3) that the company was “increasing our agility by reducing layers with fewer managers.” In May it cut about 6,000 jobs, and in July close to **9,100** more, according to [Fortune](https://fortune.com/2025/07/02/microsoft-layoffs-xbox-ceo-middle-manager-cuts/). Phil Spencer told the Xbox organization it would “follow Microsoft’s lead in removing layers of management to increase agility and effectiveness.” Meta made the argument about decisions explicit. When it cut roughly 600 roles from its AI organization in October 2025, chief AI officer Alexandr Wang wrote in an internal memo reported by [Axios](https://www.axios.com/2025/10/22/meta-superintelligence-tbd-ai-reorg): “By reducing the size of our team, fewer conversations will be required to make a decision, and each person will be more load-bearing and have more scope and impact.” Gartner had predicted the direction a year earlier. In its October 2024 list of top predictions, it said **20%** of organizations would use AI to flatten their structure through 2026, removing more than half of their middle management positions, because AI can automate scheduling, reporting and performance monitoring and so widen each manager’s span of control. The same paragraph lists the costs: managers “feeling overwhelmed with additional direct reports,” and mentoring and learning pathways that “may become broken.” Those two warnings are the subject of this note. In [an earlier note](https://vandam.ai/notes/org-chart-redrawn) I looked at the org charts Block, Coinbase and Fiverr published. This one looks at a single box on those charts: the manager. ## What got cut was the routing work Read Jassy’s complaint again and look at what he lists. Pre-meetings, reviews before a topic can move, decisions made one level away from the work. All of it is routing: carrying information up, carrying approvals down, keeping people aligned. For most of the last century that was a large share of a middle manager’s week, and it is the share software now handles best. An agent can compile a status update from the ticket system and the CRM. Microsoft’s [2025 Work Trend Index](https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born), a survey of 31,000 workers in 31 countries published in April 2025, describes where it thinks this goes. It calls the companies furthest along “Frontier Firms,” built around hybrid teams of people and agents, and predicts every employee becomes an “agent boss”: someone who builds, delegates to and manages agents. It proposes a new metric, the human-agent ratio, meaning how many agents a role needs and how many people it takes to guide them. **82%** of leaders said they were confident they would use digital labor to expand workforce capacity within 12 to 18 months, and **28%** of managers said they were considering hiring AI workforce managers to lead hybrid teams. Microsoft sells the agents, so I discount the forecast. The ratio is the useful part: it admits that agents need supervising, and someone’s week pays for it. ## Agents hand every manager a second team The best evidence on what that supervision costs comes from the company furthest ahead. In its October 2026 report [How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic), the company measured its own staff. Meta-work, meaning specs for multi-agent runs, agents reviewing other agents and pipeline monitoring, grew from **10%** of the average employee’s interactive sessions in fall 2025 to **22%** in spring 2026. Claude sat idle about 80% of the time, waiting for a person. The median employee spent **20%** more time in meetings than a year earlier, and in an April 2026 internal survey, 20% of employees answering an open-ended question described a role “shifting from actually doing work to managing Claudes.” Two findings matter most for managers. First, managers’ code commits doubled: the people supervising technical teams build and ship more themselves. Second, the authors note that agents “become teams for a single employee, rather than nodes with high connectivity to many employees at once.” A lead with ten people now oversees ten people who each run a small team of agents, and every one of those agent teams produces output that someone has to read, judge and sign. The report also shows what happens to decisions. Hard problems that agents surface end up needing people in a room, and work done in private becomes a liability because agents only know what is written down. The authors conclude that “working and deciding in the open is critical at an AI-intensive organization.” For a manager, that turns a habit into a duty: if a decision is not written where the agents can read it, the team’s agents will keep working from the old one. ## Wider spans only work under conditions Gallup can name So flattening raises the number of people per manager at the same moment agents raise the output per person. Gallup’s [January 2026 study of span of control](https://www.gallup.com/workplace/700718/span-control-optimal-team-size-managers.aspx) shows the first half already happening in the US. Average direct reports rose from **10.9** to **12.1** in a year, nearly 50% more than when Gallup first measured in 2013, pulled up by a growing minority of teams with 25 or more people. The median stays at six. The managers are paying for it. Gallup’s [2026 State of the Global Workplace](https://www.gallup.com/workplace/349484/state-of-the-global-workplace.aspx) reports that global manager engagement fell from **27%** to **22%** between 2024 and 2025, the largest one-year drop it has recorded for the group. Managers, it writes, “used to enjoy an ‘engagement premium’ at work, but they are increasingly only as engaged as those they lead.” In South Asia, where the share of managers fell, manager engagement fell eight points, and Gallup names organizational flattening as a possible factor. The span study also says when wide spans work. Manager talent matters more than team size: strong managers stay engaged with 25 or more direct reports. Individual contributor work has a ceiling: the median manager spends 40% of their time on it, and those above that line tend to run smaller teams. And the weekly conversation carries the weight: employees who strongly agree they got meaningful feedback in the past week are about seven in ten engaged regardless of team size, against one in four otherwise. Gallup says 15 to 30 minutes, held consistently, is enough. Its own summary is blunt: “increasing span of control alone is not a performance strategy.” ## A manager role, redesigned Here is how we would rewrite one role. Take a finance operations lead at a 400-person distributor. Today she has seven direct reports in accounts payable, receivables and the monthly close, plus one supervisor between her and the clerks. A large part of her week goes to status meetings, approval chains and reconciling the numbers her team reports to the controller. After the redesign the supervisor role is gone, she has eleven direct reports, and agents handle invoice matching, first-pass reconciliations and the variance commentary. Her job description gets four parts, each with a time budget and something she owns. **People, about 30% of the week.** A 20 to 30 minute conversation with each person every week, on goals, recent work and what they are learning. Hiring and promotion cases. Deciding which junior work stays human so her people still build judgment. **The agent system, about 25%.** She owns the standing instructions every agent on her team works from, each with a named author and a review date. She owns the review rules: which outputs a person signs before they leave the team, which get sampled, and what share. She reads a weekly sample herself, and she decides when an agent’s scope grows or shrinks. **Decisions, about 20%.** A written list of what she decides alone, what she escalates and what her people decide without her. Every decision she makes goes into the team’s decision log the same day, so people and agents work from the same version. **Her own build work, capped at 25%.** She improves one process a quarter herself, so she still knows what the work looks like up close. Gallup’s data suggests this share is where wide spans break, so the cap is written into the role. Three things leave the role, by name: compiling status reports, relaying approvals that her people can make, and pre-meetings. Her scorecard changes with it: close cycle time, error rate in agent output caught after sign-off, regretted attrition, and the share of her people who say they had a meaningful conversation with her that week. ## Set the span from a weekly hours budget The practical rule we use is simple arithmetic. Before widening a span, add up a manager’s week from the redesigned role. Count about an hour and a half per direct report for the weekly conversation, coaching and reviewing their work. Add the measured hours that report’s agent output takes to review and sign; this is the line most plans leave out, and Anthropic’s data says it is growing. Add six to eight hours for decisions and cross-team work, and cap build work at a quarter of the week. If the total runs past 45 hours, the span is too wide for this role, whatever the target ratio says. For the finance lead, eleven people at an hour and a half is about 16 hours. If agent review averages 40 minutes per person per week, that adds seven. Decisions take seven, build work ten. The total is about 41 hours, which fits. If agent review grows to two hours per person, the same span runs to 55 hours, and either the review rules change or the team splits. ## What to do this quarter For two weeks, have every manager in one function log their time in five buckets: routing, people, agent oversight, decisions and their own work. The routing bucket is what you can remove without removing a person. Remove the routing first. Kill the status meetings that an agent-compiled report can replace, and push the approvals Jassy described down to the people doing the work. Then rewrite one manager role per function using the four parts above, with time budgets and named artifacts, and run the hours budget before any span changes. Pick managers for the people part of the job. Gallup’s finding that talent beats team size is the strongest argument in this note, and it means the managers you keep through a flattening matter more than the number you cut. Finally, add one question to your monthly operating review: what share of each team says it had a meaningful conversation with its manager this week. If that number falls while spans widen, the flattening has gone too far. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Ask what your agents can reach because any text they read can give them orders. · VANDAM > Any text an agent reads can steer it. A one-page reach review for every agent, and a sandbox for the agents your employees already brought. URL: https://vandam.ai/notes/what-your-agents-can-reach Contact: hendrik@vibelab.be # Ask what your agents can reach because any text they read can give them orders. Prompt injection tops OWASP’s list of LLM risks, and every frontier model in a NIST-run red-teaming contest was hijacked at least once, so the limit has to sit in what an agent can touch. I lay out the one-page reach review we use and how to give employees who bring their own agents a sandbox instead of a ban. Hendrik De Winne 26 October 2026 9 min read In July 2025, researchers at [General Analysis](https://www.generalanalysis.com/blog/supabase-mcp-blog) set up a small test. A developer’s coding assistant in Cursor was connected to the company database with a service key that bypasses all row-level security. An attacker opened a support ticket and wrote instructions into it, addressed to the assistant. Later the developer asked the assistant to review open tickets. It read the message, treated the instructions as commands, pulled the table holding OAuth tokens and session credentials, and posted its contents back into the attacker’s ticket. The researchers’ summary: “No permissions were violated. The agent just followed instructions it should never have trusted.” Nobody hacked anything in that story. The assistant had broad access, it read a customer’s text, and it could write somewhere the customer could see. Everything worked as configured. The orders simply came from outside the company, because to a language model an instruction in a ticket looks much like an instruction from its user. That changes where the security decision sits. Telling an agent to behave well protects very little, since anyone who can put text in front of it can tell it something else. What protects you is what the agent can reach: which systems, which data, which ways out. That is a question a COO can ask at approval time, and it is the subject of this note. The ledger-specific controls are in [an earlier note](https://vandam.ai/notes/agents-and-internal-controls); this one covers every agent, including the ones employees brought themselves. - **13/13** Frontier models hijacked at least once in a public red-teaming competition run with [NIST’s CAISI](https://www.nist.gov/blogs/caisi-research-blog/insights-ai-agent-security-large-scale-red-teaming-competition), March 2026. - **78%** Share of AI users who bring their own AI tools to work, per the 2024 [Work Trend Index](https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part). - **58%** Share of workplace Claude use that runs through personal accounts, per [Cyberhaven](https://www.cyberhaven.com/blog/sensitive-data-flowing-into-ai-tools), February 2026. - **$50** Upper bound on the cost of the run that found a 27-year-old OpenBSD bug, per [Anthropic](https://www.anthropic.com/research/mythos-preview). ## Any text an agent reads can give it orders The security community has a name for this and ranks it first. In the 2025 [OWASP Top 10 for LLM Applications](https://genai.owasp.org/llmrisk/llm01-prompt-injection/), prompt injection is LLM01. OWASP separates the direct kind, where a user types the attack, from the indirect kind, which “occur when an LLM accepts input from external sources, such as websites or files.” Company agents read email, tickets, PDFs and web pages from strangers all day, so the indirect kind is the one that matters. The documented cases follow one pattern. In May 2025, [Invariant Labs](https://invariantlabs.ai/blog/mcp-github-vulnerability) showed that a malicious issue filed on a public GitHub repository could hijack a developer’s agent connected through GitHub’s official MCP server. The agent read the issue, went into the user’s private repositories, and published what it found, including salary information, in a pull request on the public one. In June 2025 Microsoft fixed [CVE-2025-32711](https://nvd.nist.gov/vuln/detail/CVE-2025-32711) in Microsoft 365 Copilot, rated **9.3** out of 10, which its advisory describes as “AI command injection.” The researchers who found it, at Aim Security, showed that a single email could lead Copilot to send internal data out without the recipient clicking anything. Simon Willison, the developer who coined the term prompt injection in 2022, gave the clearest test for when this becomes dangerous. In [The lethal trifecta for AI agents](https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/), published in June 2025, he names three properties: “access to your private data,” “exposure to untrusted content” and “the ability to externally communicate.” An agent with all three can be talked into reading something private and sending it to whoever asked. The Supabase and GitHub cases above each had all three. The only way to stay safe there is to avoid that lethal trifecta combination entirely. (Simon Willison, [June 2025](https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/)) ## Better models reduce the odds and leave the exposure in place The natural hope is that the next model will refuse these instructions. The evidence says it will refuse more of them, and that this is not enough to build on. In January 2025, NIST’s Center for AI Standards and Innovation [tested agent hijacking](https://www.nist.gov/news-events/news/2025/01/technical-blog-strengthening-ai-agent-hijacking-evaluations) against what was then a leading model. New attacks from its red team raised the success rate from **11%** for the strongest known attack to **81%**, and simply retrying each attack 25 times raised the average from 57% to 80%. In March 2026 CAISI reported on a [public competition](https://www.nist.gov/blogs/caisi-research-blog/insights-ai-agent-security-large-scale-red-teaming-competition) with more than 400 participants and over 250,000 attack attempts against 13 frontier models: “At least one successful attack was found against all of the target frontier models.” An attacker needs to win once and can retry almost for free. So I plan as if the injection will eventually land, and size the damage by what the agent could touch when it does. That turns an unsolved research problem into an ordinary permissions decision. ## Stolen credentials reach exactly as far One of the most cited incidents around an AI chat product did not involve prompt injection at all, although it is often retold as if it did. In August 2025, according to [Google Threat Intelligence](https://cloud.google.com/blog/topics/threat-intelligence/data-theft-salesforce-instances-via-salesloft-drift), an actor it tracks as UNC6395 used “compromised OAuth tokens associated with the Salesloft Drift third-party application” to export large volumes of data from numerous corporate Salesforce instances. It then searched the stolen records for AWS keys, passwords and Snowflake tokens. The lesson for this note is the same one, from the other side. A chat integration on marketing websites held standing tokens into its customers’ CRMs. Whether an attacker steers an agent with text or steals the keys it carries, the blast radius is set by what those keys open. Every agent and every AI integration you approve adds one more set of keys to that count. ## Your people already brought their own agents Many of the AI tools touching company data were never approved by anyone. Microsoft’s 2024 [Work Trend Index](https://www.microsoft.com/en-us/worklab/work-trend-index/ai-at-work-is-here-now-comes-the-hard-part), a survey of 31,000 workers, found that **78%** of AI users bring their own tools to work, and **80%** at small and medium companies. Cyberhaven’s [2026 report](https://www.cyberhaven.com/blog/sensitive-data-flowing-into-ai-tools), built from its customers’ data-flow telemetry, counts 32.3% of workplace ChatGPT use and 58.2% of Claude use running through personal accounts, and says **39.7%** of AI interactions involve sensitive data. Cyberhaven sells data-loss prevention, so treat these as a vendor’s view of its own customers. The bigger shift this year is that personal tools now act. OpenClaw, the open-source agent that runs around the clock and controls email, calendar and GitHub through a chat app, has passed 390,000 stars on [GitHub](https://github.com/openclaw/openclaw) since November 2025. Its project is candid about what it is built for. Its [security policy](https://github.com/openclaw/openclaw/blob/main/SECURITY.md) calls it a “personal assistant” for one trusted operator and warns that “anyone who can operate an agent can make it do anything that agent can do.” Its creator put it more bluntly. > The security model of OpenClaw is that it’s your PERSONAL assistant (one user - 1...many agents). IT IS NOT A BUS. > > Peter Steinberger, Creator of OpenClaw, February 2026, [Source](https://x.com/steipete/status/2026092642623201379) That is a reasonable design for a personal tool. Inside a company it means an employee’s agent, logged in with that employee’s credentials, can do everything the employee can, on instructions from any email that lands in their inbox. The same applies to the quick internal apps staff now build in a weekend. [Supabase’s own documentation](https://supabase.com/docs/guides/database/postgres/row-level-security), for one of the databases these apps most often sit on, warns that “a table in an exposed schema without RLS is readable and writable by any role with a grant on it.” A ban moves all of this to personal accounts, where you see none of it, and tells your most productive people their way of working is a disciplinary matter. The better move is to give them a place to run agents where the reach is decided by the company. ## A reach review takes one page per agent This is the review we use before any agent, bought, built or brought, gets access to company systems. The COO or CISO signs it. **1. Name the reach.** List every system the agent can read, every system it can write to, and every way it can send something out: email, chat, web requests, file shares, pull requests, links it renders. **2. Check for the trifecta.** Mark whether the agent reads untrusted content, holds private data and can communicate externally. If all three are true, remove one before go-live. Usually the cheapest to remove is the outbound channel: no open web requests, sending only to an allowlist, drafts that a person releases. **3. Give it its own identity.** The agent runs on its own account with scoped tokens, never an employee’s login, an admin key or a database service role. Tokens expire, and the agent appears in the access review like any user. **4. Grant per task.** Read-only by default. One project, mailbox, repository or folder per job, following OWASP’s advice to “restrict the model’s access privileges to the minimum necessary for its intended operations.” Write access is added for a named task and removed after it. **5. Put a person on the irreversible steps.** Payments, deletions, permission changes, external sends and anything touching customer records wait for a human click. Turn off “always allow” for those tools. **6. Log what it saw and did.** Keep the inputs the agent read, the tools it called and what it sent, for long enough to reconstruct an incident. **7. Attack it before launch, and again after changes.** Plant instructions in the documents, tickets and emails it will read and see what it does. Repeat after every model, prompt or connector change, and on a schedule measured in weeks. **8. Keep an off switch.** One named person can revoke the agent’s tokens in minutes, and knows which ones to revoke. The cadence in point 7 matters because finding flaws has become cheap for everyone. Anthropic reported in April that its Mythos Preview model found a 27-year-old bug in OpenBSD in a run that cost under $50, and added that the thousand runs behind it cost under $20,000 in total. In October, after AI-assisted researchers found critical flaws in PHP, the FBI advised operators who could not patch at once to take their sites offline, as the [Associated Press](https://www.adn.com/nation-world/2026/10/21/websites-scramble-to-fix-software-flaw-that-could-put-much-of-internet-at-risk/) reported. An annual penetration test and a yearly access review were designed for a slower world. ## Offer a sandbox where you would have written a ban For the agents employees bring, my rule is to say yes on the company’s terms. In practice that is five things. **A register.** A short form: which agent, which tools it connects to, who runs it. Registering earns access, so people have a reason to do it. **A separate workspace.** Agents run on a company-managed machine or cloud workspace, apart from the employee’s laptop, with their own accounts: a mailbox alias, a calendar with read access, a copy of the data they need instead of the live system, a test database instead of production. **Default scopes.** Read-only connectors to the main systems, with write access to the sandbox only. Moving an agent onto a live system goes through the reach review above. **A named voucher.** The employee answers for what their agent produces, the same way they answer for their own work. **An easy exit.** When an employee’s agent proves useful, IT helps turn it into a supported tool, which keeps the register honest. ## What to do this month For a COO or CEO, four steps. First, ask IT for a list of every OAuth grant and API token held by an AI tool or integration across email, CRM, code and file storage, and revoke the ones nobody can explain. Second, pick the three agents with the widest reach and run the one-page review on each, starting with the trifecta check. Third, set up the sandbox and the register before you write any new AI policy, so the policy has somewhere to send people. Fourth, ask your security lead or outside tester to plant instructions in a document one of your agents reads next week, and look at what it did. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Who should do your AI work, now that the labs sell it too. · VANDAM > The AI labs and investors moved into services in 2026. Four ways a mid-sized company can get its AI work done, and when each one is the right call. URL: https://vandam.ai/notes/who-does-the-work Contact: hendrik@vibelab.be # Who should do your AI work, now that the labs sell it too. In 2026 Anthropic and OpenAI set up services companies, Sequoia and Y Combinator told founders to sell the work, and Accenture bought an AI firm. A mid-sized company now chooses between a product, a big consultancy, a lab-backed team and a small senior team. Hendrik De Winne 24 October 2026 7 min read On May 4, [Anthropic](https://www.anthropic.com/news/enterprise-ai-services-company) announced a new AI services company with [Blackstone](https://www.blackstone.com/news/press/anthropic-partners-with-blackstone-hellman-friedman-and-goldman-sachs-to-launch-enterprise-ai-services-firm/), Hellman & Friedman and Goldman Sachs. The announcement says exactly who it is for: companies “from community banks to mid-sized manufacturers and regional health systems” that “lack the in-house resources to build and run frontier deployments.” Applied AI engineers from Anthropic will work inside the new firm. [Fortune](https://fortune.com/2026/05/04/anthropic-claude-consulting-industry-joint-venture-blackstone-goldman-sachs/), citing the Wall Street Journal, reported about **$1.5 billion** in committed capital. A week later [OpenAI](https://www.adventinternational.com/news/openai-launches-the-openai-deployment-company-to-help-businesses-build-around-intelligence/) launched the OpenAI Deployment Company, “majority-owned and controlled by OpenAI,” with more than **$4 billion** of initial investment from 19 investment firms, consultancies and system integrators. To staff it from day one, OpenAI agreed to buy Tomoro, an applied AI consulting firm that brings about 150 forward-deployed engineers and deployment specialists. I work with exactly the companies that sentence describes, and the two labs whose models I use every day now want to sell them people as well as software. Anthropic’s CFO, Krishna Rao, explained why in one line. Enterprise demand for Claude is significantly outpacing any single delivery model. (Krishna Rao, CFO of [Anthropic](https://www.anthropic.com/news/enterprise-ai-services-company), May 2026) Put plainly, the labs have more customers who want to use the models than people who can install them. That is an admission about where AI value gets stuck, and it comes from the company that sells the model. ## Everyone moved toward the work in the same five months The labs were not alone. In January [Accenture](https://newsroom.accenture.com/news/2026/accenture-to-acquire-faculty-to-scale-ai-capabilities) agreed to buy [Faculty](https://newsroom.accenture.com/news/2026/accenture-to-acquire-faculty-to-scale-ai-capabilities), a London firm of more than 400 AI specialists, including data scientists and AI engineers, and made its CEO, Marc Warner, Accenture’s chief technology officer. The deal closed in March. A month earlier [Accenture and Anthropic](https://www.anthropic.com/news/anthropic-accenture-partnership) had set up a joint business group in which about 30,000 Accenture people would be trained on Claude and a new role, the “reinvention deployed engineer,” would embed Claude inside client systems. In February OpenAI signed multi-year “Frontier Alliances” with BCG, McKinsey, Accenture and Capgemini, pairing its own forward-deployed engineers with the consultants, as [TechCrunch](https://techcrunch.com/2026/02/23/openai-calls-in-the-consultants-for-its-enterprise-push) reported. Forward-deployed engineer is the job title of the year. Palantir made it famous, and nearly every firm in this note now uses some version of it. Patrick Heffernan, principal analyst at [TBR](https://tbri.com/blog/consulting-will-rebound-in-2026/), expects it to spread through IT services in 2026 “primarily as a marketing term but also backed by actual organizational changes.” Both halves of that sentence are true, and a buyer has to work out which half they are being offered. - **6:1** Services spend to software spend, per [Sequoia](https://sequoiacap.com/article/services-the-new-software/), March 2026. - **$4B+** Initial investment in the [OpenAI](https://www.adventinternational.com/news/openai-launches-the-openai-deployment-company-to-help-businesses-build-around-intelligence/) Deployment Company, May 2026. - **400+** AI specialists [Accenture](https://newsroom.accenture.com/news/2026/accenture-to-acquire-faculty-to-scale-ai-capabilities) added by buying Faculty, closed March 2026. ## Investors say the money is in the work The investors explained the logic before the labs acted on it. In March [Julien Bek](https://sequoiacap.com/article/services-the-new-software/), who invests at the seed and early stage for Sequoia, published “Services: The New Software.” His example is an accounting department: a company might pay **$10K** a year for QuickBooks and **$120K** for the accountant who closes the books. “For every dollar spent on software, six are spent on services.” The founders he wants to back sell the closed books. > If you sell the tool, you’re in a race against the model. But if you sell the work, every improvement in the model makes your service faster, cheaper, and harder to compete with. > > Julien Bek, Seed and early-stage investor, Sequoia, March 2026, [Source](https://sequoiacap.com/article/services-the-new-software/) In April [Y Combinator](https://x.com/ycombinator/status/2048834289769603126) published a call for AI-native service companies. The post, by general partner Gustaf Alströmer, says it in one sentence: “We’re excited about companies that don’t sell a tool to help you do the work: they just do the work.” The part of Bek’s essay I keep coming back to is a caveat he adds near the end. He splits work into intelligence, which follows complex but learnable rules, and judgment, which “requires experience and taste.” On his map, management consulting is a “huge market but the work is mostly judgement.” Medical coding and standard NDAs sit at the other end. For a mid-sized company that tells you something specific. The narrow, rule-heavy tasks you already outsource will be sold to you as a finished product soon, if they are not already. Deciding how your company should run around AI is the judgment work, and the investors betting most heavily on AI services expect that part to stay with people. ## A mid-sized company has four ways to get the work done So a company of a few hundred people now has four serious options for its AI work. Each one is the right answer for someone. 1. **Buy a product that does the task.** If the job is narrow and countable, like drafting NDAs, coding medical bills or adjusting standard insurance claims, a vertical product that sells the outcome is usually the cheapest and fastest route. Bek’s essay is a shopping list of them. The catch: the product does the task its own way, and your process bends to fit it. It also changes nothing around the task. Your people still hand work back and forth the way they did before. 2. **Hire a large consultancy.** If you run thirty sites in twelve countries, need a regulator to sign off and want a firm with a balance sheet behind the program, this is the honest choice. Accenture now has Faculty’s engineers in-house, and McKinsey and BCG work alongside OpenAI’s. The catch is the one buyers already know: price, and the distance between the partner who designs the change and the team that builds it. 3. **Bring in a lab-backed deployment team.** If you have chosen one model and the work is mostly engineering, nobody knows that model better than the people who made it. Anthropic’s own example is a group of physician practices, where engineers sit down with clinicians and build tools around documentation, coding and prior authorizations. The catch is that one lab staffs it and that lab’s models come with it. [One post on X](https://x.com/dom__inic/status/2066260212772679864) put the longer-term risk sharply: “organizations are effectively outsourcing key learning curves and domain knowledge to the AI labs.” 4. **Work with a small senior team that redesigns and builds.** This is what I run at VANDAM. The same few people map how the work flows, decide which steps go, rewrite the roles with HR and finance, and build the agents. We are model-agnostic with Claude by preference, and the client owns what we build. The catch: we have little capacity, no global footprint and no balance sheet for a forty-country rollout. The options overlap more than the list suggests. A lab team can talk to your HR director, and a consultancy can write code. What separates them is who owns the redesign, and who is still around once the engineers leave. ## When I would send you elsewhere VANDAM is the wrong choice in at least three situations, and I would rather say so up front. If your problem is a single outsourced task with a clear output, buy the product. Paying a team to redesign a process that a vendor already sells as a finished service wastes your money and my time. If you are a portfolio company of one of the investors behind the Anthropic venture, already standardized on Claude, with a clear engineering backlog, that firm was built for you, and it has Anthropic’s own engineers inside, which I do not. If you need the same change in many countries at once, with a program office and procurement rules that only a large firm can satisfy, hire the large firm and hold it to measurable results. The same post that warned about learning curves also makes the strongest case against firms like mine: “If consulting services are truly a value add and competitive advantage, then in the long term you want to in-house.” I agree with it. The goal of an engagement should be a company that can carry on without us. That is why our work includes education, why the client keeps the code and the company context we build, and why the first deliverable is a plan your own people can argue with. Where a small senior team fits best is the company in the middle: a few hundred people, a process that crosses finance, operations and HR, nobody inside with time to redesign it, and a leadership team that wants to understand what changes before anything is built. In that company, the expensive mistake is to automate today’s process as it stands. The labs, the consultancies and the investors all agree that the work is where the value sits. Whoever you choose, ask them to name the person who will redesign it, the person who will build it, and what the people doing it today will be doing a year from now. If one name answers the first two questions, you are probably talking to a team like ours. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Who trains your next senior once agents do the junior work. · VANDAM > AI is shrinking entry-level hiring. What the data shows, why it empties the pipeline to senior roles, and how to rebuild apprenticeship around agents. URL: https://vandam.ai/notes/who-trains-the-next-senior Contact: hendrik@vibelab.be # Who trains your next senior once agents do the junior work. Four studies from the past year point the same way: AI is thinning the entry-level rung, mostly by slowing hiring. The judgment employers now look for was learned on that rung, so I set out what I would change as an employer. Hendrik De Winne 24 October 2026 6 min read Two economists at [Harvard](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5425555), Seyed Hosseini and Guy Lichtinger, followed the résumés of 65 million American workers at more than 280,000 firms to see what happens to a company’s junior staff once it starts building with generative AI. They spotted adoption through a specific kind of job ad, one that hires a “GenAI integrator.” From early 2023, junior employment at those firms fell about **9%** against non-adopting firms within six quarters. Senior employment kept its old trend. The detail I keep returning to sits in their job posting data. Within the same occupation, the tasks generative AI handles well are increasingly removed from junior job postings. The ad for the first rung is being rewritten one task at a time. And the drop came from slower hiring. Juniors already inside those firms were, if anything, leaving less often than before. The role simply stopped being posted. ## Four datasets point the same way One study is a signal. Over the past year, three more have landed on the same pattern from different data. - **9%** Share of technical hires at junior level at [Anthropic](https://www.anthropic.com/research/employment-at-anthropic), down from 27%. - **14%** Drop in the monthly job finding rate for 22 to 25 year olds in the most exposed occupations, against 2022, per [Anthropic](https://www.anthropic.com/research/labor-market-impacts). - **19%** How far employment of 22 to 25 year olds in AI-exposed jobs sits below its kept-pace path, per [Stanford](https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/). The first number comes from inside the company that makes Claude. In its October report on its own workforce, which I read closely in [an earlier note](https://vandam.ai/notes/inside-anthropic), [Anthropic](https://www.anthropic.com/research/employment-at-anthropic) reports that the share of juniors in its technical departments slipped from **21%** to **14%** since early 2025. The proportion of junior technical hires “fell even faster, from 27% to 9%.” By April, engineering and research had deemphasized hiring for lower band technical roles. The second comes from Maxim Massenkoff and Peter McCrory, economists at Anthropic, in [Labor market impacts of AI](https://www.anthropic.com/research/labor-market-impacts), published in March. Using US survey data, they found no rise in unemployment for exposed workers. What they did find was that young workers became less likely to be hired into the most exposed jobs: a **14%** drop in the job finding rate compared with 2022, with “no such decrease for workers older than 25.” They are honest that the result is “just barely statistically significant,” and that the young people who were not hired may simply have stayed in the jobs they had. The third is the largest. Erik Brynjolfsson, director of the [Stanford Digital Economy Lab](https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/), and his Stanford coauthors Bharat Chandar and Ruyu Chen work from ADP payroll records covering millions of US workers. In their August 2026 update, employment of 22 to 25 year olds in AI-exposed occupations stands **19%** below where it would be had it kept pace with less-exposed peers. Experienced workers show no comparable gap. The divergence has “widened steadily” since they first published it a year earlier, and it runs mainly through reduced hiring. They call their findings early, descriptive indicators, and they flag that the pattern weakens when they control for education. I take that caution seriously. Four teams, four datasets and one direction is still hard to wave away. ## The seniors everyone wants were made on that rung Look at what employers now say they want. Anthropic’s hiring managers “increasingly emphasize applicants’ autonomy and judgment relative to specific technical expertise.” The Stanford team finds the same split in the wider market: young workers lose ground in occupations built on codified knowledge, while occupations that rely on tacit knowledge see faster employment growth for experienced workers. The market pays for the things that are hard to write down. Those things were learned on the bottom rung. Hosseini and Lichtinger describe junior work as “intellectually mundane tasks,” routine but cognitively demanding, such as debugging code or reviewing legal documents. A junior did the dull version of the task next to someone who knew what good looked like, got corrected, and slowly absorbed why. That loop is where judgment came from. It was never written into a job description, because the work itself carried it. [Enrique Ide](https://arxiv.org/abs/2507.16078), an assistant professor of economics at IESE Business School, built a model of exactly this loop. In it, novices acquire tacit knowledge by working alongside experts, and growth comes from the best experts passing their practice on to several novices at once. His conclusion is sharper than a headcount story. > I show that improvements in entry-level automation increase output upon adoption but can reduce growth and welfare, even without reducing entry-level employment. > > Enrique Ide, Assistant Professor of Economics, IESE Business School, June 2026, [Source](https://arxiv.org/abs/2507.16078) The damage in his model comes from moving novices away from the most productive experts. So keeping the same number of juniors only helps if they still learn from the strongest people and still get to touch the real work. ## The bill arrives in five years, at HR Every individual decision here makes sense. A team lead with a good agent setup does not need a new graduate this year, and the business case for skipping the hire writes itself: one salary saved, no onboarding, no mistakes to fix. Nobody books the cost, because it lands years later in a different budget. Around 2031, the company looks for people ready to run a close, lead a client or own a system, and finds a gap where its 2026 intake should have been. The obvious answer is to hire those people from the market. That works for one company. When every company reasons the same way, the market has no one left to sell, because nobody trained them. For a single employer it is a pipeline question. Across an industry it becomes a commons that everyone draws on and nobody refills. The strongest objection deserves a fair hearing. Many juniors cost more than they produced in their first year, and some of that early work really was filler that an agent should do. The research also cuts the other way: experimental studies find generative AI raises productivity most for less experienced workers, as Hosseini and Lichtinger point out. I agree with both points, and they lead somewhere useful. A junior with good agents is worth more than a junior without them. That is an argument for redesigning the role, and a weak argument for deleting it. ## Four changes I would make as an employer **Keep the entry path on purpose.** Put a number in the workforce plan: how many entry-level people you take on each year per senior, reviewed once a year like any other capacity. If the number goes to zero, that should be a decision someone signed, with the five-year consequence written next to it. **Rebuild the apprenticeship around supervising agents.** The junior runs the agent on the work they would once have done by hand: the reconciliation, the first draft of the contract review, the bug triage. Their job is to check the output, find what is wrong and explain why. A senior then reviews the junior’s review. The junior stays close to the object-level work, which is where judgment forms, and the senior’s correction lands on reasoning instead of typing. Following Ide, I would pair juniors with the strongest people on the team, even though those are the people with the least spare time. **Measure judgment directly.** If hiring managers now score judgment at the door, track it after the door too. Give juniors agent output with errors you already know about and record how many they catch. Log the calls they escalate and check later whether escalating was right. Within a few months you have a picture of who is ready for more, based on evidence a manager can defend. **Count teaching as work.** The pairing costs senior hours. If those hours are invisible in objectives and reviews, they disappear in the first busy quarter. I would write a teaching load into senior roles and treat it like any other deliverable. This is the part of training I care about most in the [100xworker Program](https://100xworker.com), because it is the part companies skip when they focus on tools. A first step costs an afternoon. Pull your hiring data for the last three years, count entry-level hires per senior in each team, and check whether the ratio fell. If it did, find out whether anyone decided that, or whether it happened one skipped requisition at a time. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Why most AI pilots stall and how to choose one that scales. · VANDAM > What MIT's 95% AI failure figure really measured, why pilots stall before production, and a five-point rule for choosing a pilot that scales. URL: https://vandam.ai/notes/why-pilots-stall Contact: hendrik@vibelab.be # Why most AI pilots stall and how to choose one that scales. MIT's famous 95% figure measures less than the headline says, but larger surveys from S&P Global, Gartner and McKinsey point the same way. The pilots that reach production share five conditions, and a CEO can check for them before the first demo. Hendrik De Winne 25 October 2026 8 min read A COO at a mid-market manufacturer gave MIT researchers the most honest summary of corporate AI I have read. “The hype on LinkedIn says everything has changed, but in our operations, nothing fundamental has shifted. We’re processing some contracts faster, but that’s all that has changed.” The quote sits on page five of [The GenAI Divide](https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf), the July 2025 report from MIT’s Project NANDA that gave the world its favorite AI statistic: 95% of organizations getting zero return. That number traveled further than the report did. [Fortune](https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/) ran it as 95% of generative AI pilots failing, and it has been on conference slides ever since. I think the pattern behind it is real. The number itself needs a careful reading, and once you read it carefully, the more useful question is what the surviving pilots had in common. - **5%** Organizations whose task-specific gen AI tools reached production with sustained impact, per [MIT NANDA](https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf), 2025. - **42%** Companies abandoning most of their AI initiatives before production, up from 17%, per [S&P Global](https://www.spglobal.com/market-intelligence/en/news-insights/research/ai-experiences-rapid-adoption-but-with-mixed-outcomes-highlights-from-vote-ai-machine-learning). - **46%** Average share of projects scrapped between proof of concept and broad adoption, same survey. - **6%** Share of companies getting at least 5% of EBIT from AI, flat for a year, per [McKinsey](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai), 2026. ## The 95% figure measures less than the headline says The MIT team drew on three sources: a review of more than 300 publicly disclosed AI initiatives, structured interviews with people from **52** organizations, and a survey of **153** senior leaders collected at four industry conferences, all between January and June 2025. That is a small, self-selected sample, and the authors say so. Their own limitations section warns that organizations willing to talk about AI may differ from those that declined, and that a six-month observation window may understate success for complex systems. The definition of success matters more. A task-specific tool counted as “successfully implemented” when users or executives remarked that it caused “a marked and sustained productivity and/or P&L impact.” That is a high bar measured by recollection. Next to the key chart, the authors call the figures “directionally accurate based on individual interviews rather than official company reporting.” The funnel behind the number also tells a different story from the headline. Of the organizations studied, **60%** evaluated custom or vendor-sold AI tools, **20%** reached a pilot, and **5%** reached production. My arithmetic on their numbers: roughly one in four organizations that ran a pilot got it into production with lasting impact. Most of the loss happens before anyone pilots anything. For general-purpose tools such as ChatGPT and Copilot, the report puts pilot-to-implementation at around 83%, though it adds that those tools mostly help individuals and rarely show up in the P&L. So the honest reading is narrower. In a small 2025 sample, very few companies could point to custom AI that changed their numbers within months. That is still a bad result. It just measures something different from 95% of pilots failing. ## Larger surveys land in the same place The reason I still take the pattern seriously is that bigger samples with different methods agree. In its 2025 Voice of the Enterprise survey of **1,006** IT and line-of-business managers in North America and Europe, [S&P Global Market Intelligence](https://www.spglobal.com/market-intelligence/en/news-insights/research/ai-experiences-rapid-adoption-but-with-mixed-outcomes-highlights-from-vote-ai-machine-learning) found that the share of companies abandoning most of their AI initiatives before production rose from **17%** to **42%** in one year. On average, respondents scrapped **46%** of projects between proof of concept and broad adoption. Gartner saw it coming. In July 2024 it predicted that at least **30%** of gen AI projects would be [abandoned after proof of concept](https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025) by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value. In June 2025 it added that over **40%** of [agentic AI projects will be canceled](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) by the end of 2027, and estimated that only about 130 of the thousands of vendors selling agentic AI offer the real thing. These are forecasts, so treat them as informed expectation. The 2026 data has not changed the picture much. In [McKinsey’s State of AI 2026](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai), fielded in May and June with 1,719 respondents, **37%** say AI has contributed to EBIT at all, essentially unchanged from 2025, even as more companies report scaling. The share of high performers stayed at about **6%**. More deployment has not yet turned into more profit for most companies. ## Most pilots are picked because they demo well The first structural reason is how pilots get chosen. A pilot has to win approval, and approval goes to what looks impressive in a meeting. MIT’s respondents put about half of their hypothetical AI budget into sales and marketing, and the authors explain why: outcomes there are easy to measure and map to board-level KPIs. The back office, where they found faster payback from cutting outsourcing and agency spend, gets less money because its gains are harder to show. One CIO in the report described the result: “We’ve seen dozens of demos this year. Maybe one or two are genuinely useful.” A demo proves the model can do the task once, on clean inputs, in front of an audience. Production asks whether it can do the task ten thousand times inside a process that people already run. A pilot chosen for the first test tells you little about the second. ## Nobody in the business owns the result The second reason is ownership. Many pilots are run by an innovation team, an IT group or a vendor, and handed to the business at the end. The business never asked for it, so the business has no reason to change anything to make it work. MIT found that the strongest deployments were sourced from frontline managers and budget holders, with executive accountability on top. Lack of executive sponsorship was among the top barriers its respondents named. The buy-versus-build data points the same way once you look closely. MIT reports that pilots built with external partners reached deployment about **67%** of the time against about **33%** for internal builds, while noting the figures are self-reported and may reflect differences between the companies. In September 2026, [Gartner](https://www.gartner.com/en/newsroom/press-releases/2026-09-29-gartner-predicts-70-percent-of-enterprises-will-abandon-agentic-ai-built-by-vendor-forward-deployed-engineering-by-2028) predicted that by 2028, **70%** of enterprises will abandon agentic AI built by vendor forward-deployed engineers, “trapped by soaring costs and unable to evolve it on their own.” I read the two together: outside help can speed up the first deployment, and the company still has to own what gets built. > Identify an executive sponsor who is accountable for business outcomes, not simply the implementation budget. > > Mukul Saha, Sr Director Analyst, Gartner, September 2026, [Source](https://www.gartner.com/en/newsroom/press-releases/2026-09-29-gartner-predicts-70-percent-of-enterprises-will-abandon-agentic-ai-built-by-vendor-forward-deployed-engineering-by-2028) ## The money stops when the pilot ends The third reason is budget. Pilots are often paid from an innovation or IT discretionary line with a fixed end date. Running the thing in production needs a different line: licenses at full volume, integration work, monitoring, someone to fix it when the vendor changes a model. If nobody approved that line before the pilot started, a successful pilot arrives at budget season as a new cost with no owner, competing against everything else. Gartner’s Rita Sallam made the related point in 2024: gen AI costs “aren’t as predictable as other technologies,” and many CFOs have not been comfortable investing today for indirect value later. A pilot that cannot show its full running cost and its payback in the owner’s own budget will usually lose that argument. ## The data was good enough for a demo The fourth reason is data. A pilot runs on a sample someone cleaned by hand. Production runs on whatever sits in the ERP, the CRM and the shared drive. In a July 2024 survey of 1,203 data management leaders, [Gartner](https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk) found that **63%** of organizations either lack or are unsure they have the right data management practices for AI. It predicts that through 2026, organizations will abandon **60%** of AI projects unsupported by AI-ready data. The gap rarely shows during the pilot, and usually surfaces in the first month of real volume. ## The process around the tool stayed the same The fifth reason is the one I see most often. The company adds AI to a step and leaves the rest of the process untouched: the same approvals, the same handoffs, the same people checking the same things. The time saved disappears into the queue at the next step. MIT’s explanation for failed tools was “brittle workflows, lack of contextual learning, and misalignment with day-to-day operations.” McKinsey’s 2026 high performers do the opposite. They “fundamentally redesign workflows that are enabled by AI rather than insert AI into existing ones,” and they are twice as likely as others to report that senior leaders show commitment and that the company has defined processes to measure impact. Gartner’s agentic AI release makes the same recommendation: in many cases, rethinking the workflow from the ground up is the better path. ## Choose a pilot that has an owner before it has a demo This is the rule we use. Before a pilot starts, five things must be true. If any one is missing, fix it first or pick a different pilot. **1. A named owner who will run it.** One person in the business, with the P&L or the KPI it will move, who agrees in writing to operate it after the pilot. An IT lead or a vendor does not count. **2. A run budget approved up front.** The full annual cost at production volume, in the owner’s budget, signed before kickoff. The pilot money and the run money are approved in the same meeting. **3. A baseline measured today.** One process metric, such as cycle time, cost per case, error rate or outside spend, measured on current work before anything changes. Without it, nobody can show the result later. **4. Real data, tested early.** In week one, run the tool on a hundred real cases pulled from production systems, unedited. If it fails there, you have learned what the pilot would have taught you in month six. **5. The new process written down.** One page that says which steps disappear, who reviews what, and what the people freed up will do next. If nobody can write that page, the pilot will save time that nobody captures. Add a fixed decision date. MIT found that the top mid-market performers took about 90 days from pilot to full implementation, against nine months or longer for large enterprises. I set the date at 90 days: scale it, fix one named problem, or stop. Within those rules, I would favor back-office work with outside spend attached, such as outsourced document processing or agency fees, because the saving is visible in an invoice. ## What to do this quarter Ask for a list of every AI pilot running in the company, with its owner, its cost so far and its intended production date. Put the five questions next to each one. Many companies will find that few pilots have a business owner and fewer still have a run budget. Then make three calls. Stop the pilots that fail more than two of the questions, and say so openly so the teams can move on. Give the rest 30 days to fill the gaps, with the COO checking. And approve no new pilot until its owner can answer all five in one meeting, with the CFO in the room for the second one. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # The work you never did because it never paid is where the largest AI return sits. · VANDAM > The largest AI return is the expert work a mid-market firm skips today. What Jevons, ATMs and translation show, and how to pick the first three. URL: https://vandam.ai/notes/work-you-never-did Contact: hendrik@vibelab.be # The work you never did because it never paid is where the largest AI return sits. Most AI business cases count the hours of work a company already does, which caps the return at today's labor cost. The evidence from Jevons to bank tellers says cheaper work expands where demand was rationed, and I set out how a leadership team lists, prices and picks the expert work it skips. Hendrik De Winne 26 October 2026 8 min read Last spring, Austin Lau, then a growth marketer at Anthropic, described a Google Ads workflow he had built with Claude. Once a company spends enough on search, he wrote, “no marketer is ever going to go through tens of thousands of search terms daily, let alone EVERY single search term to identify negative keywords.” So the agent does it, every day. In [his own summary](https://x.com/helloitsaustin/status/2041218668072403356), it “was building a workflow for something I’d never do.” Of the four ways he used AI, he called this one “the one that most people overlook.” That is the gap in almost every AI business case I read. The shortlist starts from the work people do today, and the saving is capped at what that work costs today. If your company spends 400 hours a year on contract review, the best automation can ever return is 400 hours. The review you skip on 80% of your deals has no hours attached, so it never appears on the shortlist, even though the money it touches can be far larger than the hours. Economics has a precise answer for when cheaper work turns into more work, and it is useful to a CFO. It depends on whether demand was being held back by price. Demand for a mid-market company’s own expert work is held back about as hard as demand gets, because the company rations it: contract review on the big deals, churn analysis on the top accounts, a pricing review once a year. Rationed demand is exactly the kind that expands when the price falls. - **2013** Tellers needed per urban bank branch, 1988 to 2004, as ATMs spread, per [James Bessen](https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm) (2015). - **43%** Growth in urban bank branches over the same years, so teller jobs did not disappear. - **61%** Share of New York-area service firms using AI in 2026, up from 25% in 2024, per the [New York Fed](https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/). - **4%** Share of those service firms that laid off workers in response to AI in the past six months. ## Cheaper work has usually meant more of it The idea is old. In 1865 the English economist William Stanley Jevons argued in [The Coal Question](https://energyhistory.yale.edu/files/2023/06/Jevons-The-Coal-Question.pdf) that more efficient steam engines would raise Britain’s coal use. “It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth.” Cheaper coal per unit of work made coal worth using in places it had never paid. Economists now call this the Jevons paradox. The service version I find most useful is the bank teller. ATMs spread through American banking from the mid-1990s, and the obvious forecast was that tellers would vanish. [James Bessen](https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm), an economist at Boston University, showed in 2015 that it went the other way. The number of tellers needed to run an average urban branch fell from **20** to **13** between 1988 and 2004. A branch became cheaper to run, so banks opened more of them, and urban branches grew by **43%**. “Fewer tellers were required for each branch, but more branches meant that teller jobs did not disappear.” Aaron Levie, CEO of Box, makes the same case for marketing in his essay [Jevons Paradox for Knowledge Work](https://x.com/levie/status/2004654686629163154). He estimates that the US had a few hundred thousand people in marketing-related jobs in the 1970s and has low millions today, because software let almost any small business “justify the ROI of doing more sophisticated marketing.” He calls his figure back of the envelope math, and he runs a company that sells AI to businesses, so I read it as his estimate. The direction matches the teller story. ## It fails where demand was already full The paradox has a condition, and it decides where your money should go. In a [2019 paper in Economic Policy](https://ideas.repec.org/a/oup/ecpoli/v34y2019i100p589-626..html), Bessen traced two centuries of US textile, steel and auto employment. Automation first added jobs in all three, then cut them. In his words, “What changed? The elasticity of demand.” While people wanted much more cloth than they could afford, cheaper cloth meant more cloth and more workers. Once people had enough, each new machine simply removed labor. Translation shows the failure case today. Carl Benedikt Frey and Pedro Llanos-Paredes of [Oxford](https://www.oxfordmartin.ox.ac.uk/publications/lost-in-translation-artificial-intelligence-and-the-demand-for-foreign-language-skills) studied 695 US local labor markets after Google Translate launched as an app, and published in January 2025. Areas that adopted machine translation more heavily saw translator employment fall, and employer demand for foreign language skills fell with it. In that market the cheap version substituted for paid work instead of adding to it. The rule I take from the two cases: falling prices create new work where demand was rationed by price, and remove work where demand was already met. That turns the question for a leadership team into a concrete one: which expert work does it ration today because the full version costs too much. ## Mid-market companies ration expert work the hardest A company of 400 people rations expert work all the time without calling it that. Legal reads the master agreements on the large deals and lets the rest go out on the template. Customer success studies the accounts that are already shouting. Pricing gets reviewed when someone has a week free. Invoices are audited by sample. Win and loss interviews happen after the deals people remember. Each of those is a decision that the full version is not worth an analyst’s salary. A Fortune 500 company makes far fewer of those trade-offs, and Levie puts the difference plainly in the same essay. > For most large companies today, they can effortlessly move resources around between projects, afford to experiment on new ideas, hire the top lawyers or marketers for any new project they need > > Aaron Levie, CEO, Box, December 2025, [Source](https://x.com/levie/status/2004654686629163154) His claim is that “AI agents bring democratization to every form of non-deterministic knowledge work,” and his list starts with reviewing contracts, market research and 24/7 support. I covered his budget argument in [the note on pricing AI per finished task](https://vandam.ai/notes/cost-of-the-i). The part that matters here is where he points: to the work a smaller company has never been able to afford. In April he gave the operating test in [a post on agent roles](https://x.com/levie/status/2043883641366032638): look for tasks where agents let you “either execute it 100X faster or do it 100X more times than before.” Most shortlists only look for the first. ## So far, companies are reorganizing more than cutting The early data fits this picture better than the layoff headlines do. In the New York Fed’s [regional business survey](https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/), published in September,, **61%** of service firms used AI, yet only **4%** had laid off workers because of it in the previous six months. About 15% said they had hired fewer people than they otherwise would, and 13% said they had hired more. Among AI users, just over a third of service firms were retraining staff. The authors conclude that “existing workers are much more likely to be retrained than replaced by AI.” Denmark offers the deepest look so far. Anders Humlum and Emilie Vestergaard linked chatbot adoption surveys to national employment records in an [NBER working paper](https://www.nber.org/papers/w33777) revised this March. They find “precise null effects on earnings and recorded hours” two years after ChatGPT, while “new AI-related tasks are widespread,” including content generation, AI oversight and AI integration. Employers are absorbing the technology by changing the work. None of that happens on its own. [Nikita Bier](https://x.com/nikitabier/status/2004985707870445855), former head of product at X, noticed in December that once the cost of making software and video fell, “all the same people are doing it.” Cheaper tools changed what was possible and left the question of who starts untouched. The same holds inside a company: the skipped contract review stays skipped until someone decides it is now worth doing, gives it an owner and a budget, and acts on what it finds. ## How to list, price and pick the work you skip This is the exercise we use. It takes one leadership meeting and a week of follow-up. **1. List.** Each executive brings five items, answering three prompts. What do we check by sample that we could check in full? What do we do yearly or quarterly that would pay monthly or weekly? What does a company twice our size do that we skip? Typical answers: contract review on every deal, a churn read on every account every month, a monthly pricing and discount review, a full audit of vendor invoices, win and loss interviews on every closed deal, competitor tracking on every product line. **2. Write down the decision.** For each item, name the decision its output would change and the person who makes that decision. An item with no decision attached is a report, and it comes off the list. **3. Price it twice.** First, what it would cost with people at today’s rates. That number explains why you skip it. Second, what it costs with agents: the run cost plus the human hours to review the output and act on it. The review hours are usually the larger line, and they are the one most cases leave out. **4. Size the money it touches.** Estimate the revenue, leakage or exposure the item reaches: the renewal base the churn read covers, the spend the invoice audit covers, the contract value that goes out unreviewed. Then make a conservative guess at the share you would act on. Write the guess down so you can test it later. **5. Pick three.** Rank by money touched times your conservative share, divided by the agent-era cost. Take the top item from three different categories, ideally one about revenue, one about leakage and one about risk, so a single bad assumption cannot sink the whole program. My rule for the shortlist: an item makes it only if a named person has agreed to act on its output within a week of receiving it, and the output is something a person can check and mark as wrong. Run each of the three for 60 days alongside current practice. The test is simple: what did the full version find that the sampled version missed, and what was it worth? At the end you will know whether your demand for that work was rationed, which is the only condition under which the Jevons logic pays. ## What to do this month If you run the company, put the exercise on the next leadership agenda and ask every function head for their five items in writing beforehand. If you run finance, add a second section to the AI business case template, next to hours saved, titled work we will now do, with the decision it feeds, the owner, and the money it touches. If you run operations, take the first item on the list that runs on data you already have, such as invoice lines or support tickets, and run the full version once by hand with an agent this week, so the leadership meeting starts from a real result. ## More notes 29 October 2026 6 min read ## Who owns what your AI learns. Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections. 26 October 2026 9 min read ## Delete the step first and give the agent whatever survives. Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go. ## Your agents need a critic and so does the plan they carry out. Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it. --- # Personal AI Setup by VANDAM: Your own assistant, set up properly > One week to set up your personal AI properly. Connected to your mail, calendar, documents and CRM with the permissions you choose. Your priorities, people and writing voice written down. A handful of skills for the work that fills your week URL: https://vandam.ai/personal-ai-setup Contact: hendrik@vibelab.be # An assistant that knows your week and writes the way you do. One week to set up your personal AI properly. Connected to your mail, calendar, documents and CRM with the permissions you choose. Your priorities, people and writing voice written down. A handful of skills for the work that fills your week, and coaching until it becomes a habit. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Personal%20AI%20Setup) **$4,800** one week · per person ## The week goes to meetings and the mail around them. - **72%** of a CEO’s work time goes to meetings, 37 in an average week. Porter and Nohria, 27 CEOs, [Harvard Business Review](https://hbr.org/2018/07/how-ceos-manage-time), 2018. - **117** emails reach the average worker every day, most of them skimmed in under a minute. [Microsoft](https://www.microsoft.com/en-us/worklab/work-trend-index/breaking-down-infinite-workday), Work Trend Index, June 2025. - **2 hours** less a week on email for workers who used an AI assistant built into their mail, meeting and writing apps. Field experiment with 7,137 workers in 66 firms, [NBER](https://www.nber.org/papers/w33795), 2025. > The most valuable skill sets on the planet right now: 1. people who can set up agents properly, manage them, and run local AI models > > Greg Isenberg, CEO, Late Checkout, June 2026, [Source](https://x.com/gregisenberg/status/2068325550603559402) Most of the hour before a meeting goes to finding the last email. ## Most of an executive’s week is meetings and the mail around them. When Michael Porter and Nitin Nohria tracked how 27 CEOs spent their time, meetings took 72% of it and electronic communication another 24%, reported in [Harvard Business Review](https://hbr.org/2018/07/how-ceos-manage-time). The meetings are the job. The hours around them, sorting the inbox, reading back through a thread before a call, turning a week of notes into an update for the board, are work an assistant can do well once it can see what you see. That last condition is where most setups stop. An executive with a chat window and no access to mail, calendar or CRM gets general answers, and pasting context in by hand costs the time it was meant to save. Setting it up properly means a short list of decisions: what it may read, what it may draft, what it may never send, who your people are and how you write. The setup takes a week because those decisions are yours, and we make them with you, in your own accounts. You leave able to change every one of them yourself. > I’m starting to believe more and more that the most effective employees will have custom agents and personal software they bring to their jobs. > > Cody Schneider, Graphed, November 2025, [Source](https://x.com/codyschneider/status/1993817048896426147) ## How the week runs. 1. Connect what you work in: Mail, calendar, documents and CRM are connected on day one, with your IT lead where needed. It reads by default and drafts when asked. Sending stays with you unless you decide otherwise for a specific kind of message. 2. Write your context down: A session of about ninety minutes on your priorities this quarter, the people who matter and how you work with each of them, and the way you write, checked against your sent mail. The result is a short file you can read and edit. 3. Build your skills, then use them: Inbox triage, meeting briefs from the CRM and past mail, board and investor updates and decision memos, adjusted to how you work. You use them on your real week with a short daily check-in, and a follow-up a month later. ## Set up in a week. Adjusted as your work changes. Priorities shift every quarter, people change roles and new kinds of work turn up. After the setup we can check in each quarter to refresh your context, tune the skills you use and add the ones you now need. - Context refresh: Priorities, people and projects updated for the new quarter. - New skills: The next repeat task turned into a skill. - Access check: Connections and permissions reviewed against what you agreed. - Coaching call: An hour on what you use, what you skip and why. [Ask about quarterly check-ins](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Personal%20AI%20Setup%20care) ## What you receive. - Connected assistant: Reads what you allow, drafts and waits for you to send. - Context file: Priorities, people and writing voice in a file you can edit. - Personal skills: Inbox triage, meeting briefs, updates and decision memos. $4,800 one week · per person An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For founders, CEOs and executives at 20–500 person companies. Useful when the week goes to email, meeting prep and updates. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Personal%20AI%20Setup) ## Where the setup leads. Each next step has its own service. ### Team Harness From one person to the whole team, with the same care. **$12,800** ### Company Brain Company knowledge your assistant can answer from, with the sources attached. **$19,200** ### AI Colleague A weekly call on what changed and what to add to your setup. **$2,400/mo** ## Before you ask. ### Which assistant do you set up? Usually Claude, connected to Google Workspace or Microsoft 365 and your CRM. If your company already works with another assistant, we set up that one. ### Can it send mail on my behalf? Only if you decide it may. By default it drafts and you send. You can allow sending for one kind of message, such as meeting confirmations, and keep everything else for yourself. ### What about confidential mail? We agree with you, and with your IT lead where needed, which accounts, folders and labels it may read. Board and HR folders often stay out at first. ### How much of my time does the week take? About four hours: a kick-off, the context session and short daily check-ins while you use it on real work. ### Can my executive assistant be involved? Yes, and it often works best that way. Your assistant or chief of staff joins the context session and learns to keep the skills and context up to date. --- # Pilot Workshop by VANDAM: Design one AI pilot with your leadership team in a day > One day with the CEO, CFO, head of HR and your legal lead. We start from a result you want to move, pick the workflow closest to it, mark every step to eliminate, automate or delegate, check what is ready and agree when the pilot goes on or URL: https://vandam.ai/pilot-workshop Contact: hendrik@vibelab.be # Design your first AI pilot in one day with your leadership team. One day with the CEO, CFO, head of HR and your legal lead. We start from a result you want to move, pick the workflow closest to it, mark every step to eliminate, automate or delegate, check what is ready and agree when the pilot goes on or stops. You leave with a one-page pilot card the team can start on the week after. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Pilot%20Workshop) **$4,800** one day ## Many companies start AI pilots. Far fewer carry them into production. - **42%** of companies abandoned most of their AI initiatives before production, up from 17% a year earlier. [S&P Global](https://www.ciodive.com/news/AI-project-fail-data-SPGlobal/742590/) survey of more than 1,000 respondents in North America and Europe, 2025. - **46%** of AI proofs of concept were scrapped before they reached production at the average company, in the same survey. - **63%** of finance organizations said AI implementation was slower than expected in 2025. [Gartner](https://cpapracticeadvisor.com/2026/05/28/gartner-says-cfos-must-stop-mistaking-finance-ai-deployment-for-value-creation/184094) survey of 204 finance leaders, March 2026. > Before deciding what work/product to transform, you need to decide the outcome. > > Alex Lieberman, Co-founder, Tenex, September 2026, [Source](https://x.com/businessbarista/status/2101659025788031152) Bring whoever has to sign off, so the pilot is decided in the room. ## Most pilots start from a tool someone saw in a demo. Yours starts from a number you want to move. A pilot usually begins with a tool. Nobody writes down which result it should move, who owns that result or what would make the company stop, and a few months later it is still running without anyone able to say whether it worked. Many companies start pilots and far fewer carry them into production, as our [notes on why pilots stall](https://vandam.ai/notes/why-pilots-stall) lay out. The workshop runs the other way around. The leadership team agrees on one outcome in a plain form: improve this measure from this baseline to this target by this date. Then we take the workflow closest to it, draw it as it runs today and mark every step. E means eliminate it, A means an agent automates it, D means delegate it to the person who should own it. > During the workshop, we asked every exec to write out one of their/their teams key processes step by step and place an E (eliminate), A (automate), or D (delegate) next to each step. > > Alex Lieberman, Co-founder, Tenex, April 2026, [Source](https://x.com/businessbarista/status/2047488697885413391) Finance, HR and legal sit at the table for the steps that touch money, people and risk, so their questions get answered on the day and come back as conditions on the pilot. The day ends with a readiness check across eight areas and the rules for going on or stopping. Once the pilot runs, an [AI Opportunity Map](https://vandam.ai/ai-opportunity-map) ranks what comes next, and a [Company Redesign](https://vandam.ai/company-redesign) takes on the whole company. ## How the day runs. 1. Morning: the outcome and the workflow: We agree one outcome with a baseline, a target and a date, list the workflows close to it and pick the one that would move it most. Then we draw that workflow as it runs today, step by step, with who does each step and how long it takes. 2. Early afternoon: mark every step: Each step gets a letter: E for the steps that can go, A for the ones an agent can take, D for the ones that belong with someone else. The new process is drawn next to the old one, with the time each step should take. 3. Late afternoon: readiness and the card: We rate eight areas from value at stake to legal controls, decide what has to be fixed first and write the criteria to go on or stop. The pilot card is signed off before everyone leaves. ## Designed in a day. Built when you are ready. Some teams run the pilot themselves from the card. Others ask us to build it and to review the results with them against the go and stop criteria, on the dates written on the card. - Pilot build: The agent and connections for the marked process, built on your stack. - Go or stop review: A session at each decision date, against the criteria on the card. - Measure tracking: The measure followed from its baseline, so the result is clear on the day. - Next workflow: When the pilot passes, the next workflow designed the same way. [Ask about building the pilot](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Pilot%20Workshop%20care) ## What you receive. - Pilot card: One page with the outcome, owner, scope, measure and go/stop rules. - Marked process: The workflow today and redesigned, every step marked. - Readiness check: Eight areas rated, with what to fix before the pilot starts. $4,800 one day An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For leadership teams at 20–500 person companies. Useful when AI is on the agenda and nobody has picked the first thing to try. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Pilot%20Workshop) ## Where the workshop leads. Each next step has its own service. ### Floor Plan Comes before the day: the workflows to choose from, already drawn with their hours. **$6,400** ### Agent-Ready Audit Checks the data and systems the pilot depends on before it starts. **$9,600** ### Team Harness Rolls out the tools to the pilot team, with the rules agreed on the day. **$12,800** ### Eval Suite Measures the pilot against the go and stop criteria on your own cases. ### AI Opportunity Map Once the pilot runs: what to start next, ranked on the same terms. ## Before you ask. ### Who needs to be in the room? The CEO, the CFO, the head of HR and whoever covers legal, plus the owner of the workflow if that is someone else. Up to eight people works well. ### Do we need to prepare anything? A short call the week before to agree two or three candidate outcomes. A baseline number for each makes the morning go faster. ### What if we pick the wrong pilot? The stop criteria are written so a wrong pick shows early. The card says what has to be true by which date, and what happens if it is not. ### How does this relate to the AI Opportunity Map? The workshop designs one pilot in one day. The map looks across the whole company and ranks where to start. Once the pilot runs, the map shows what comes after it. --- # Private AI by VANDAM: Run sensitive steps on a model you control > We map which steps of a workflow touch data that should stay with you, run an open-weight model on your hardware or an EU host for those steps, and route the rest to a frontier model such as Claude. Tests on your own cases show where the lo URL: https://vandam.ai/private-ai Contact: hendrik@vibelab.be # Sensitive steps on your own model and Claude for the rest. We map which steps of a workflow touch data that should stay with you, run an open-weight model on your hardware or an EU host for those steps, and route the rest to a frontier model such as Claude. Tests on your own cases show where the local model is good enough. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Private%20AI) **$16,000** three weeks · plus hardware or hosting at cost ## Privacy holds back AI in Europe. Nearly half the companies that considered it say so. - **49%** of EU companies that considered AI and did not use it named concerns about data protection and privacy. [Eurostat](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises), 2025. - **53%** named unclear legal consequences as a reason, in the same [Eurostat](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) survey. - **20%** of EU enterprises used any AI technology in 2025. [Eurostat](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises). > Dell for private AI (AI in a box for firms that won’t use the cloud) > > Greg Isenberg, CEO, Late Checkout, October 2026, [Source](https://x.com/gregisenberg/status/2113721373482987970) Most companies we meet are fine with the cloud for most of it. Just not for the sick notes. ## Most of a workflow can use the cloud. A few steps should stay home. A sick note, an employment contract, a client file in a dispute. Some data should not go to an outside model, because of the law, a works council, a client contract or plain judgment, and many AI projects stop right there. Yet in most workflows only a few steps touch that data. Reading the doctor’s note is one of them. Checking what the collective agreement allows is a different kind of step. Open-weight models can now run those steps on a server you control. [Aleph Alpha](https://x.com/Aleph__Alpha/status/2106306840657297814) released Kolibri in October 2026, built in Europe, with weights under Apache 2.0 to run on your own hardware. A month earlier [Perplexity](https://x.com/perplexity_ai/status/2094803515264978953) let its Computer start a task in the cloud and move the steps with private files to a local model on a Mac. [Y Combinator](https://www.ycombinator.com/rfs#own-your-intelligence) asks for startups that help companies own their intelligence, and for [AI for the air-gapped economy](https://www.ycombinator.com/rfs#ai-for-the-air-gapped-economy). We split the workflow by the data each step touches. Sensitive steps run on an open-weight model in your building or with an EU host. Names and identifiers are removed before anything goes out, and the remaining steps go to a frontier model such as Claude, which still does the hardest reasoning best. Every step is tested on your own cases with both models, so the split rests on measured scores. The model, the routing and the tests are yours, as described in [own your intelligence](https://vandam.ai/notes/own-your-intelligence) and on [our models page](https://vandam.ai/models). In June 2026 [@ryiacy](https://x.com/ryiacy/status/2066260212772679864) described where companies that hand their AI work to outsiders end up: paying for the training of the model that replaces them, unless they “find a way to build and own their own model ecosystem.” Running the sensitive steps yourself is a practical first part of that. ## How the split gets made. 1. Map the data in each step: We walk through the workflows with the people who run them and with your lawyer or data protection officer. Each step is marked by the data it reads and writes: health, personnel, client or public. The map says which steps stay inside and why. 2. Install the local model: We choose an open-weight model that fits the job and your hardware, and run it on a server in your office or with an EU host. Routing sends each step to the right model, and names and identifiers are removed before text goes out. 3. Test both, then hand over: Every step runs on your own cases with the local model and with a frontier model such as Claude. The scores show where local is good enough. You get the setup, the routing rules, the tests and a runbook for updates, outages and new workflows. ## Set up once. More steps move inside as models improve. Open-weight models get better every few months. After the handover we can test each new release on your cases, update the local model when it clears the bar and move more steps inside as the scores allow. - Hosting: The local model kept running, patched and monitored, on your hardware or with the EU host. - Model updates: New open-weight releases tested on your cases before they replace the current one. - Routing review: The routing rules and the log of what left the building checked every month. - Review: A monthly session with IT, HR and legal on what moved inside and what comes next. [Ask about model care](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Private%20AI%20care) ## What you receive. - Sensitivity map: Every step, the data it touches and where it may run. - Hybrid setup: A local model, routing rules and a log of what left the building. - Eval comparison: Scores per step for the local and the frontier model, plus a runbook. $16,000 three weeks · plus hardware or hosting at cost An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For companies of 20 to 500 people with workflows whose data should stay in the building. Useful when legal, a works council or a client contract stops an AI project over where the data goes. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Private%20AI) ## Where Private AI leads. Each next step has its own service. ### Eval Suite Proves the local model does the work, on your own cases, before you switch. **$12,800** ### Token Ledger What moved in-house saves, and what still runs outside. **$9,600 + $1,600/mo** ### Agent Ops We run the local models and the agents on them after go-live. **$3,200/mo** ## Before you ask. ### What hardware do we need? For one or two workflows, usually one server with a recent GPU, in your office or rented from an EU host. We size it on your volumes in the first week. Hardware or hosting is billed at cost. ### Why not run everything locally? Frontier models such as Claude still reason better on the hard steps, and most steps in a workflow touch nothing sensitive. The tests show per step where a local model is good enough. ### Does this make us GDPR compliant? It keeps health and personnel data on infrastructure you control and logs what goes where. Your lawyer or data protection officer makes the legal call, with the sensitivity map and the log in hand. ### Which open-weight model do you use? The one that scores best on your cases and fits your hardware. We test a few, European models among them, and you can switch later without rebuilding the routing. ### Who owns the setup? You do. The model runs on your hardware or your account with the host, and the routing rules, tests and runbook are yours. --- # Source Cleanup by VANDAM: Clean the data your agents will read > Three weeks to merge duplicate records, archive dead sources and agree what your key terms mean. Your agents read one customer, one product list and one definition of churn, and every merged field shows the rule that decided it. URL: https://vandam.ai/source-cleanup Contact: hendrik@vibelab.be # Give your agents one version of every customer. Three weeks to merge duplicate records, archive dead sources and agree what your key terms mean. Your agents read one customer, one product list and one definition of churn, and every merged field shows the rule that decided it. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Source%20Cleanup) **$12,800** three weeks ## Company data holds several answers. Agents pick one. - **26%** of their organization’s data is untrustworthy, data and analytics leaders estimate. [Salesforce](https://www.salesforce.com/news/stories/data-analytics-trends-2026/) State of Data and Analytics, November 2025. - **19%** of company data is siloed, inaccessible or otherwise unusable, by the same leaders’ estimate. 70% of them think their most valuable insights sit in that 19%. [Salesforce](https://www.salesforce.com/news/stories/data-analytics-trends-2026/), 2025. - **60%** of AI projects without AI-ready data will be abandoned through 2026, [Gartner](https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk) predicts. > Marie Kondo for your company’s data (clean it up so agents can use it) > > Greg Isenberg, CEO, Late Checkout, October 2026, [Source](https://x.com/gregisenberg/status/2113721373482987970) ## Agents repeat whatever they find. Most companies keep several answers to the same question. Point an agent at your systems and it reads every version of a customer with the same confidence. The CRM has one address, the ERP another, a spreadsheet from 2023 a third. The agent picks one, and the answer sounds right whichever it picked. Data and analytics leaders estimate that 26% of their organization’s data is untrustworthy, according to [Salesforce](https://www.salesforce.com/news/stories/data-analytics-trends-2026/). The [Agent-Ready Audit](https://vandam.ai/agent-ready-audit) shows where this blocks the first agent. Source Cleanup fixes it. We match the duplicates for the two or three entities your agents need most, settle each conflicting field with a rule the field’s owner signs off, archive what nobody maintains and write down what words like active customer, ARR and churn mean at your company. [Company Brain](https://vandam.ai/company-brain) and every agent after it builds on the result. > How do you deal with fragmented data? What is your process to unify without a massive overhaul of our orgs data architecture? > > Alex Lieberman, Co-founder, Tenex, July 2026, [Source](https://x.com/businessbarista/status/2077143133049995619) Your systems stay where they are. We clean the records inside them, keep a match table that links every source ID to its one record and give agents a read-only connection to the result. A company of 40 needs the same work at a smaller scale, which is why the scope is two or three entities and three weeks. ## How the three weeks run. 1. Match the records: We pull customers, products or suppliers from every system that holds them and match duplicates on VAT number, address, email domain and name. Your team confirms the matches we are unsure of. 2. Settle the conflicts: For each field we agree a rule with its owner: legal name from the ERP, contact from the most recent email, status from the last invoice. The rules run on every record and stay written down next to it. 3. Retire, define, connect: Sources nobody maintains are archived with a pointer to what replaced them. Definitions are agreed with the teams that use them. The clean records and the glossary are exposed to your agents through a read-only MCP connection. ## Cleaned once. Kept clean every month. New records arrive with the old habits. After the cleanup we can run the matching and the field rules every month, send new duplicates to their owners and keep the glossary current. - Duplicate sweep: New duplicates matched with the same rules, uncertain ones sent to their owner. - Rule review: Field rules adjusted when a system or a process changes. - Source check: New channels, folders and files reviewed before agents read them. - Glossary session: A monthly check with the owners on definitions that need to change. [Ask about monthly upkeep](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Source%20Cleanup%20care) ## What you receive. - Golden records: One record per customer, product or supplier, with the rule behind every field. - Source register: What stays, what is archived and who owns each source. - Shared definitions: A glossary your teams and agents read from. $12,800 three weeks An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For operations, finance and IT leads at 20–500 person companies about to connect agents to company data. Useful after an Agent-Ready Audit, or when a pilot keeps answering from the wrong record. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Source%20Cleanup) ## Where the cleanup leads. Each next step has its own service. ### Agent-Ready Audit Comes before: finds which sources need cleaning for the agents you plan. **$9,600** ### Company Brain Reads the clean sources and keeps them current. **$19,200** ### Agent Onboarding Connects agents to one clean version of the data. **$9,600 + $1,600/mo** ### Invoice to ERP Matches invoices against supplier and article data that now agrees. **$24,000** ## Before you ask. ### Which entities do you clean? Two or three, chosen with you at the start. Usually customers plus products or suppliers, because those are the records your first agents read most. ### Do we need to replace our CRM or ERP? No. Records are cleaned where they live, and a match table links every source ID to its one record, so your systems keep working as they do today. ### What happens to the sources you archive? They move to an archive with a pointer to what replaced them, and agents stop reading them. Nothing is deleted without the owner’s approval. ### Who decides which value wins? The owner of each field. We propose a rule, show the records it would change and run it once they agree. --- # Team · VANDAM > The people behind VANDAM, and the operations, people, finance, legal and engineering seats each project combines. URL: https://vandam.ai/team Contact: hendrik@vibelab.be # Redesigning a company takes more than engineers. AI changes roles, budgets and how decisions get made. So every VANDAM project combines operations, people, finance and engineering, with the people below and specialists we bring in per project. ## The people behind it. Two brothers, two co-founders: one leads the AI and the story, the other the numbers and the strategy. - Hendrik De Winne: Co-founder · AI and marketing Founded the Ghent AI studio VibeLab in 2024 and has led AI projects from the first workshop to the handover since. Claude Ambassador, member of Anthropic’s partner program, and the host of Belgium’s first Claude Code event. [LinkedIn](https://www.linkedin.com/in/hendrik-de-winne-88883a192) - Thomas De Winne, CFA: Co-founder · Finance and strategy CFA charterholder and finance consultant in reporting and cash flow forecasting. Worked in corporate development and business intelligence at Telenet. Studied economics at Ghent University and at Rotterdam School of Management. [LinkedIn](https://www.linkedin.com/in/thomas-de-winne-cfa-108a52152) ## The seats we fill per project. Each project gets the disciplines it needs. These are the people we are bringing in: each has done the work in their own field, at the companies we work for. [Take a seat](mailto:hendrik@vibelab.be?subject=VANDAM%20%2F%20Joining%20the%20team) - Open seat: Leadership Former CEO of a 100 to 500 person company Sits with your leadership team and decides what changes first, and who carries it. - Open seat: Technology Former CIO of a multi-site business Checks systems, access and security before anything connects to your data. - Open seat: Operations Former COO or head of operations Maps how the work flows today and redesigns the steps around agents. - Open seat: People and HR Former CHRO or HR director Redesigns roles, brings employees in early and plans who learns what. - Open seat: Legal and compliance Former general counsel Reads the contracts, the data rules and the audit trail before anything goes live. - Open seat: Change and training Former head of learning and development Turns a new way of working into a habit, team by team. --- # Team Harness by VANDAM: Roll out Claude to a whole team, safely > Claude Code and Claude Cowork, set up for your team in three weeks. Permissions and approved connectors agreed with IT, a shared library of skills written from your own SOPs, each with an owner, a link to your company context and a usage st URL: https://vandam.ai/team-harness Contact: hendrik@vibelab.be # Give the whole team what your power users built for themselves. Claude Code and Claude Cowork, set up for your team in three weeks. Permissions and approved connectors agreed with IT, a shared library of skills written from your own SOPs, each with an owner, a link to your company context and a usage standard people can follow. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Team%20Harness) **$12,800** three weeks ## Leaders use AI every day. The front line has stalled. - **51%** of frontline employees use AI regularly, a figure that has stalled, while more than three quarters of leaders and managers use it several times a week. [BCG](https://www.bcg.com/publications/2025/ai-at-work-momentum-builds-but-gaps-remain), AI at Work, 10,600 people, June 2025. - **1/3** of employees say they have been properly trained to use AI. [BCG](https://www.bcg.com/publications/2025/ai-at-work-momentum-builds-but-gaps-remain), AI at Work, 2025. - **78%** of AI users bring their own AI tools to work. [Microsoft and LinkedIn](https://news.microsoft.com/source/2024/05/08/microsoft-and-linkedin-release-the-2024-work-trend-index-on-the-state-of-ai-at-work/), Work Trend Index, 31,000 people, 2024. > How do I ensure that I roll out claude code and cowork securely without putting my company data at risk? > > Alex Lieberman, Co-founder, Tenex, July 2026, [Source](https://x.com/businessbarista/status/2077143133049995619) Start with the SOP people actually follow. The one in the shared drive is often from 2021. ## Most teams have a few power users and a lot of logins. In most companies we work with, three or four people have built a serious setup of their own: saved prompts, personal skills, connections to the tools they use every day. Everyone else has an account and uses it like a search box. [Alex Lieberman](https://x.com/businessbarista/status/2056879209583349801), a founder of [Tenex](https://tenex.co), heard the same from the CHRO of a $500m consumer business and wrote it down as their key challenge: “A widening gap between power users and the rest of the org; reluctance to roll advanced access to all employees without guardrails is creating tension.” Both halves of that sentence are reasonable. Claude Code and Cowork read files, run commands and act in connected tools, so somebody has to decide what each role and each skill may touch. Once that is decided, the methods of the few can become the team’s. A skill is a written procedure the assistant follows, and most of yours already exist as SOPs, onboarding docs and the checklists your best people keep. The harder part is keeping it shared. [J.B.](https://x.com/VibeMarketer_/status/2086808854898323774) describes what most teams end up with: “Each person has to teach their AI what they know and how they work. The context, corrections, and workflows they create rarely reach anyone else.” A library with an owner per skill and a review step fixes that. One person’s improvement becomes everyone’s starting point the day the owner approves it. > A skill is a scary word. It’s nothing more than an SOP. > > Alex Lieberman, Co-founder, Tenex, April 2026, [Source](https://x.com/businessbarista/status/2047488697885413391) ## How the team gets set up in three weeks. 1. Set the boundaries first: With IT and the team leads we decide who gets Claude Code, who gets Cowork, which connectors are approved and what each may read, draft or change. Data that stays out is written down. Accounts, roles and settings are configured to match, on your plan and your identity provider. 2. Write the first skills from your SOPs: We collect the procedures the team already follows and pick the 10 to 15 that run most often. Each becomes a skill with a named owner, the connectors it may use and the steps it must never take. The owner reviews it against real cases before it is published. 3. Roll it out team by team: The library goes live with a distribution path, so every person has the same skills and updates arrive without copying files around. The harness links to your company context, such as a Company Brain. Each team gets a working session on its own skills, plus the usage standard. ## Set up in three weeks. Kept current every month. Models change, connectors get added and teams write better versions of their skills. After the rollout we can look after the library each month: review the changes owners approved, retire skills nobody runs and add the next procedures worth sharing. - New skills: The next procedures turned into skills, reviewed by their owners. - Library review: Skills tested again after a model or connector change. - Access check: Connectors and permissions compared with the agreed list. - Usage report: Which skills ran, for whom, and which nobody uses. [Ask about monthly care](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Team%20Harness%20care) ## What you receive. - Configured harness: Claude Code and Cowork with permissions and connectors per role. - Skills library: 10 to 15 reviewed skills, each with a named owner. - Usage standard: What goes in, what stays out and how work gets checked. $12,800 three weeks An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For COOs, CTOs and team leads at 20–500 person companies. Useful when a few people work with AI every day and the rest have a login. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Team%20Harness) ## Where Team Harness leads. Each next step has its own service. ### 100×worker Program Training on each person’s own work, so the rollout turns into results. **$8,000** ### Personal AI Setup A fuller setup for the founders and executives who live in their inbox. **$4,800** ### Company Brain Shared company context every skill and session can read. **$19,200** ### Token Ledger Usage and spend per person and team, next to what it produced. **$9,600 + $1,600/mo** ## Before you ask. ### Does it have to be Claude? We usually suggest Claude Code and Cowork, because they run skills and connectors the way this setup needs. If your team has chosen another harness, the library, owners and usage standard are built the same way on that one. ### Which SOPs do you need from us? Whatever the team follows today: procedure documents, onboarding guides, checklists, a good example email. Where nothing is written, we interview the person who does the work and write it with them. ### Who owns a skill after you leave? A named person in your team, usually the one who owned the procedure. They approve changes, and the library shows who that is for every skill. ### How do you keep company data safe? Access, connectors and data boundaries are agreed with your IT lead before anyone gets a login. Each skill lists the connectors it may use and the actions it must never take, and those limits are set in the configuration as well as in the text. ### How many people can you set up? The price covers one team of up to about forty people. Larger rollouts reuse the same library and standard and are scoped per team. --- # Token Ledger by VANDAM: AI spend per person, team and model, next to what it produced > A ledger of every AI subscription and API key in your company, with spend per person, team, tool, model and workflow next to the work it produced: tickets closed, quotes sent, hours saved. We set it up on your own accounts, then run the cap URL: https://vandam.ai/token-ledger Contact: hendrik@vibelab.be # See where the tokens go and what they produce. A ledger of every AI subscription and API key in your company, with spend per person, team, tool, model and workflow next to the work it produced: tickets closed, quotes sent, hours saved. We set it up on your own accounts, then run the caps, routing rules and monthly review with you. [Discuss the project](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Token%20Ledger) **$9,600 + $1,600/mo** two weeks · then monthly ## Company AI spend tripled in a year. Few can say what it bought. - **$37B** spent by companies on generative AI in 2025, up from $11.5 billion in 2024. [Menlo Ventures](https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/), December 2025. - **$12.5B** of it went to foundation model APIs, the usage-billed layer underneath most AI tools. [Menlo Ventures](https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/), December 2025. - **40%+** of agentic AI projects will be canceled by the end of 2027, [Gartner](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) predicts, due to escalating costs, unclear business value or inadequate risk controls. June 2025. > Historically, leaders had to think about allocating budget for headcount and software. Now, token budget must be a serious consideration. Do you set limits per employee? How do you measure the ROI on your token budget? > > Alex Lieberman, Co-founder, Tenex, April 2026, [Source](https://x.com/businessbarista/status/2047488697885413391) Most of the bill is fine. We go looking for the part that produced nothing. ## The AI bill has no line items. Nobody can say what it bought. Most companies now pay for AI in four or five places at once: chat subscriptions per seat, API keys on a developer’s card, agent platforms billed by usage, coding tools billed by the token. Finance sees the invoices. Nobody sees which team spent what on which model, or whether the work got done. Token Ledger puts every subscription and key on one page and attributes each dollar to a person, a team, a model and a workflow. Then it follows the dollar into the systems where the work lands, so the ledger shows the tickets closed, the quotes sent and the invoices posted. You see the hours saved, the cost per completed task and the share of spend that produced nothing anyone can find. That second view sets the controls. Routine work moves to a cheaper model, the frontier models stay where they change the result, and caps go on the sessions that run unattended. Two of the questions [Alex Lieberman](https://x.com/businessbarista/status/2077143133049995619)’s team hears most often from enterprise executives: “How do I start controlling token spend and how do I think about attributing value to a token?” and “How do we manage costs, and prevent runaway sessions?” > Token costs will become a dominant topic in enterprises going forward with AI. Just got out of a dinner with many Fortune 500 enterprise CIOs and this was the most heated topic. > > Aaron Levie, CEO, Box, May 2026, [Source](https://x.com/levie/status/2056965292753146019) ## How the ledger is set up and run. 1. Connect every account: We list every AI subscription, API key and agent platform in use, including the ones paid on personal cards, and connect their billing and usage data with read-only access. The ledger runs on your accounts and stores its data with you. 2. Tie spend to work: Each session is attributed to a person, team, model and workflow. We link workflows to the systems where their output lands, your helpdesk, ERP, CRM or code repository, so a dollar can be followed to a closed ticket or a sent quote. 3. Set the controls: With your finance and IT leads we set spend caps, access tiers by role, routing rules that send routine work to a cheaper model and alerts on sessions that run away. Every month we review the numbers, re-route what should move and report to leadership. ## Set up in two weeks. Reviewed every month. Models change price, new tools get adopted and teams find new uses for the ones they have. Each month we review the ledger with your finance and IT leads, move work to another model where the numbers say so, adjust caps and tiers, and send a one-page report to leadership. - Monthly report: Spend, output, hours saved and cost per task, next to last month. - Re-routing: Workflows moved between models when price or quality changes. - Caps and alerts: Caps, tiers and runaway alerts tuned to how each team works. - New accounts: New subscriptions, keys and agent platforms added as they appear. [Ask about the monthly review](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Token%20Ledger%20care) ## What you receive. - Spend ledger: Every subscription and key, by person, team, tool, model and workflow. - Output view: Tasks completed, hours saved and cost per completed task, per workflow. - Controls: Spend caps, access tiers, model routing and alerts on runaway sessions. - Monthly review: Re-routing decisions and a one-page report for leadership. $9,600 + $1,600/mo two weeks · then monthly An indication from past projects. The final price can move a little with the size and complexity of your company, and we agree it in writing before we start. For finance, IT and operations leaders at 50–500 person companies. Useful once the AI bill grows faster than anyone can explain it. [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20/%20Token%20Ledger) ## Where the ledger leads. Each next step has its own service. ### Eval Suite Finds the cheaper model that does the same work, tested on your own cases. **$12,800** ### Private AI Moves sensitive or high-volume steps to a model you control. **$16,000** ### AI Opportunity Map Moves spend to where the hours are, ranked on the same terms. **$9,600** ### Agent Ops We run the agents the ledger watches, and fix what it flags. **$3,200/mo** > A dead simple app that sits between you and all your AI agents and tracks what they’re spending. Token costs, API calls, subscriptions. One dashboard. Your AI budget is about to become a real line item. > > Greg Isenberg, CEO, Late Checkout, April 2026, [Source](https://x.com/gregisenberg/status/2048381278953603406) ## Before you ask. ### Where does the ledger run? On your accounts. It reads billing and usage with read-only keys and keeps its data in your environment. If you stop working with us, the ledger and its history stay with you. ### Which tools can it read? We check what each of your tools exposes in the first week. Most model providers offer usage and cost data through an admin API. Where a tool only sends invoices and seat lists, we attribute from those, and personal-card spend comes in through expense claims. ### How do you measure hours saved? Per workflow, with the people who do the work: time per task before and after, checked against a sample of real tasks. Each figure shows how it was measured, so finance can question it. ### Do caps slow people down? Caps are set per team and per workflow with the team leads, and most of them never trigger. The ones that matter sit on unattended sessions, where an agent in a loop keeps spending while nobody watches. ### Which models do you recommend? We are model-agnostic and start from what you already use, usually with Claude among them. Routing rules follow your own cost per completed task, measured in the ledger. --- # VANDAM Weekly · VANDAM > VANDAM Weekly: a free weekly email from Hendrik De Winne on what got released in AI, what it means for companies of 20 to 500 people, and what to try or ignore. URL: https://vandam.ai/weekly Contact: hendrik@vibelab.be # VANDAM Weekly. What changed in AI this week. A short email from Hendrik once a week: what got released, what it means for companies of 20 to 500 people, one thing worth trying and one worth ignoring. Free, and readable in five minutes. [Subscribe by email](mailto:hendrik@vibelab.be?subject=Subscribe%20to%20VANDAM%20Weekly&body=Please%20add%20me%20to%20VANDAM%20Weekly.) ## Who writes it Hendrik De Winne reads the releases every week for the companies VANDAM works with: the frontier labs, the open-source models, the AI features inside Microsoft 365, Google Workspace and the other tools teams already pay for, and the rules coming out of Brussels and Washington. The Weekly is the general version of that reading. The longer arguments go into the [notes](https://vandam.ai/notes). ## For your own company The Weekly is written for everyone, so it cannot know your stack, your industry or the projects you have open. The [AI Colleague](https://vandam.ai/ai-colleague) does the same reading against your company and brings it to a weekly call. - **VANDAM Weekly** One email for every reader What changed in AI this week Free - **AI Colleague** A weekly call of 30 to 60 minutes, with a brief sent before it What changed for your company, your industry and the work you have running A monthly price, stop any month ## Recent notes - [**Who owns what your AI learns.** Y Combinator now asks for startups that help companies own their intelligence. Before a mid-sized company owns a model, it can own the four things every model works with: context, evals, rules and connections.](https://vandam.ai/notes/own-your-intelligence) - [**Delete the step first and give the agent whatever survives.** Musk’s five-step method puts automation last, and most AI programs start there because an agent is the one step a vendor can sell. I walk accounts payable through the steps in his order and show why an agent that clears exceptions fast hides the steps that should go.](https://vandam.ai/notes/delete-the-step-first) - [**Your agents need a critic and so does the plan they carry out.** Models grade their own work too kindly, and the labs now separate the agent that makes from the agent that checks, the way finance separates maker and checker. I set out a four-layer review design with a minutes budget, and why the brief itself needs a critic with no stake in it.](https://vandam.ai/notes/agents-need-a-critic) **VANDAM Weekly** Once a week, free. Send one email to subscribe and one to stop. [Subscribe](mailto:hendrik@vibelab.be?subject=Subscribe%20to%20VANDAM%20Weekly&body=Please%20add%20me%20to%20VANDAM%20Weekly.) --- # Why VANDAM · An AI transformation studio run from Belgium > VANDAM is an AI transformation studio run from Ghent, Belgium, for companies in Europe and the US: published prices, redesign before build, and you keep everything. URL: https://vandam.ai/why-vandam Contact: hendrik@vibelab.be # A New York name, with Belgian roots. VANDAM is an AI transformation studio that redesigns how work gets done around AI agents, then builds the systems and hands them to your team, for companies of 20 to 500 people in Europe and the United States. Hendrik De Winne Ghent, Belgium ## What VANDAM does We start with how the work moves through your company today, decide what should change, and then build the systems that carry the new way of working. Every engagement has a published price and ends with your team running what we built. ### Strategy Every project starts with a [Floor Plan](https://vandam.ai/floor-plan): the people, roles and workflows of your company as they work today, credited when you continue. The [AI Opportunity Map](https://vandam.ai/ai-opportunity-map) ranks your AI ideas by hours saved and readiness, the [Agent-Ready Audit](https://vandam.ai/agent-ready-audit) checks whether your data and systems can support agents, and a [Pilot Workshop](https://vandam.ai/pilot-workshop) designs one pilot with your leadership team in a day. Each ends with a plan you can act on. ### Implementation [Company Brain](https://vandam.ai/company-brain), [Invoice to ERP](https://vandam.ai/invoice-to-erp), [Agent Onboarding](https://vandam.ai/agent-onboarding), [Account Brains](https://vandam.ai/account-brains) and [Team Harness](https://vandam.ai/team-harness) bring AI into the work your team already does. They are handed over with the code, the documentation and the training to maintain them. ### Engineering [MCP Retrofit](https://vandam.ai/mcp-retrofit), [Eval Suite](https://vandam.ai/eval-suite) and [Private AI](https://vandam.ai/private-ai) are software built around your business, and the [Adversarial App Audit](https://vandam.ai/adversarial-app-audit), [Agent Attack Audit](https://vandam.ai/agent-attack-audit) and [Local Law 144 Audit](https://vandam.ai/local-law-144-audit) are independent audits with verified findings and the fixes attached. ### Education The [100×worker Program](https://vandam.ai/100x-worker-program) assesses and trains your people on their own work, the [Judgment Vault](https://vandam.ai/judgment-vault) documents how your experts decide, and [AI Colleague](https://vandam.ai/ai-colleague) keeps leadership current every week. ## What makes VANDAM different ### The price is on the page Every service lists its price and its term, from $4,800 to $24,000, before anyone gets on a call. Many consultancies only put a number on the table after a paid discovery phase. ### Redesign comes before the build Most AI projects speed up the process a company already has. We map where the hours go first and work with HR and finance alongside engineering, which is why our [team hires from both](https://vandam.ai/notes/ai-needs-hr-and-finance). ### You keep everything The code, the prompts and the documentation belong to you, in files your company controls. Our note on [vendor exits](https://vandam.ai/notes/plan-for-your-ai-vendor-leaving) explains why that matters when a provider changes its terms. ### Model-agnostic, with Claude depth We choose models around your work. Hendrik is a Claude Ambassador and a member of Anthropic’s partner program, and hosted Belgium’s first Claude Code event, with 100 people in the room and 600 applications. ### Built under European rules Working from Belgium means GDPR and the AI Act shape the design from the first week: who can see what, who signs off and what gets logged. A US finance team asks for the same controls once an agent can [post to the ledger](https://vandam.ai/notes/agents-and-internal-controls). ## Who uses VANDAM - Leadership teams at 20 to 250 person companies deciding which AI projects to start - Operations and finance teams at multi-site businesses with invoices, orders and approvals to automate - Owners and department heads whose key decisions depend on one person - B2B sales teams at 20 to 100 person companies researching target accounts - Product and engineering teams at B2B software companies connecting their product to AI assistants - HR, people and legal teams using automated hiring tools in New York City ## The team behind VANDAM VANDAM was founded by two brothers. Hendrik De Winne founded the Ghent AI studio [VibeLab](https://vibelab.be/en) in 2024 and has run discovery, architecture, builds and handovers for retail, social-profit and drinks companies since. He also runs [100×worker](https://100xworker.com), a training program on working with background AI agents, and VibeLab Afterhours for builders in Ghent. Thomas De Winne, CFA, is a finance consultant in reporting and cash flow forecasting who worked in corporate development and business intelligence at Telenet. He sets the baseline before a project and reports what changed. VANDAM grew out of that work in 2026, with a New York name and a European base; meet everyone on the [team page](https://vandam.ai/team). Find Hendrik on [LinkedIn](https://www.linkedin.com/in/hendrik-de-winne-88883a192) and [GitHub](https://github.com/henddwin). 1624 Walloon families sail west to found New Netherland. Today VANDAM works from Ghent with companies in Europe and the United States. Europe is early in this. [Eurostat](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) counted **20%** of EU companies with ten or more staff using AI in 2025, with **55%** of large firms against **17%** of small ones. Most of the gain is still ahead, which is the argument of our note on [Europe’s diffusion bet](https://vandam.ai/notes/europe-diffusion-bet). ## How VANDAM works You start with an email to Hendrik and a call about what you want to change. We agree on the service, its price and its term in writing, and Hendrik leads the work from the first call to the handover. Engagements run one to four weeks. You see working versions along the way, and every build ends with a handover session, the documentation and training for the people who will run it. The [approach pages](https://vandam.ai/approach/strategy) describe each phase. ## Where the name comes from Vandam Street runs three blocks through Hudson Square, the old printing district on the west side of Manhattan. We took the studio’s name from the building on its corner with Greenwich Street, and the story of that ground starts in the Low Countries. ### 1624 1924 half dollar for the founding of New Netherland, U.S. Mint. Public domain. Walloon families from the southern Netherlands, a region that is largely Belgium today, sailed on the Eendracht and the Nieu Nederlandt to settle New Netherland, the colony that became New York, according to the [New Amsterdam History Center](https://newamsterdamhistorycenter.org/2023/08/29/the-first-families/). ### 1731 Rip Van Dam, portrait via Wikimedia Commons. Public domain. [Rip Van Dam](https://en.wikipedia.org/wiki/Rip_Van_Dam), a merchant born to a Dutch family in Beverwyck, now Albany, became acting governor of New York. Henry Moscow’s (Street Book) names Vandam Street after him; Sanna Feirstein’s (Naming New York) credits a liquor dealer, Anthony Van Dam, as [Forgotten New York](https://forgotten-ny.com/?p=46197) notes. ### 1797 Richmond Hill House, between Charlton and Vandam Streets. C. H. Haswell, 1897. Public domain. Aaron Burr owned the Richmond Hill estate here and mapped its land into plots of **25 by 100 feet** on three streets, which became Charlton, King and Vandam, according to the city’s [Landmarks Commission](https://s-media.nyc.gov/agencies/lpc/lp/0224.pdf). ### 1888 The 1888 brick at 100 Vandam. Photo: Kidfly182, CC BY-SA 4.0, cropped and dithered. A six-story Romanesque Revival building in load-bearing brick went up at Vandam and Greenwich as a coal-fired power plant, according to [CityRealty](https://www.cityrealty.com/nyc/market-insight/features/future-nyc/100-vandam-garden-rung-condos-west-soho-prepare-sales-launch/13061). ### 2023 The tower inside the walls. Photo: Kidfly182, CC BY-SA 4.0, cropped and dithered. [COOKFOX](https://cookfox.com/projects/100-vandam/) kept the brick walls, took out the interior and built a **25-story** tower inside the old façade. That building is the method. When factories got electric power, most swapped the steam engine for an electric motor and saw little change; the gains came once they rebuilt the factory around the new power source, as Paul David showed in (The Dynamo and the Computer) (1990). AI is at the same point. - **94%** Tasks in computer and math jobs that language models could take on, per [Anthropic](https://www.anthropic.com/research/labor-market-impacts), March 2026. - **33%** Share of those tasks Claude covers in practice. We keep the walls, replace the core, and build higher than the old thing could. The walls are your customers, your people and the judgment they built up. The core is how work moves through the company and who decides what. ## Key facts - **Company** VANDAM - **Type** AI transformation studio - **Founded** 2026, out of VibeLab (founded 2024) - **Founders** Hendrik De Winne · [LinkedIn](https://www.linkedin.com/in/hendrik-de-winne-88883a192) Thomas De Winne, CFA · [LinkedIn](https://www.linkedin.com/in/thomas-de-winne-cfa-108a52152) · [Team](https://vandam.ai/team) - **Headquarters** Ghent, Belgium - **Markets** Europe and the United States - **Website** [vandam.ai](https://vandam.ai) - **Core offering** Redesigning how work gets done around AI agents, then building and handing over the systems - **Pricing** Published prices from $4,800 for a one-day Pilot Workshop, as indications from past projects; the final price follows the size and complexity of the company and is agreed before we start - **Contract terms** Price and term agreed in writing before we start, from one day to ten weeks or more per engagement - **Services** [Twenty-four services](https://vandam.ai/services), from a one-day Pilot Workshop to systems built into the work - **Models** Model-agnostic: Claude, ChatGPT, Gemini, Copilot, Grok and open-source models, chosen per task - **Communication** Email, directly with the founder: [hendrik@vibelab.be](mailto:hendrik@vibelab.be?subject=VANDAM%20%2F%20Let%E2%80%99s%20talk) - **Alternatives** Large consultancies, AI development agencies and the AI labs’ own services teams ## Frequently asked questions ### Where is VANDAM based? VANDAM is run from Ghent, Belgium. We work with companies in Europe and in the United States, and the name comes from Vandam Street in New York. ### Which AI models do you use? We are model-agnostic and choose models around your work, your data and the tools your team already uses. Claude is our preference, and Hendrik is a Claude Ambassador and a member of Anthropic’s partner program. ### What does a project cost? Every service has a published price and term, from $4,800 for a one-day Pilot Workshop. The prices are indications from past projects: the final number can move a little with the size and complexity of your company, and we agree it in writing before we start. A Company Redesign combines several services and is scoped to the size of the company. ### Who owns what you build? You do. Each build ends with the code, the documentation and the training your team needs to run and change it without us. ### Do you use our data to train models? No. We work through the providers’ commercial terms; Anthropic, for example, does not train on commercial inputs and outputs by default. ### How do we start? Send an email to Hendrik with what you want to change. You can start with an AI Opportunity Map or go straight to the service that fits the problem. [Talk to Hendrik](mailto:hendrik@vibelab.be?subject=VANDAM%20%2F%20Let%E2%80%99s%20talk)