Check every invoice, not one in twenty, now that a judgment costs a tenth of a cent. · VANDAM

This page as Markdown llms.txt llms-full.txt services.json sitemap.xml
# 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.