I ran this weekly audit at a $180M SaaS business last quarter. It caught two revenue recognition errors that would have blown a Q3 covenant test, one duplicate vendor payment ($47,200 to a marketing agency we paid twice in 11 days), and a stock comp accrual that was $310,000 light because someone forgot a new-hire grant from a July 21 board consent.

Total time investment: 45 minutes on a Friday. The controller runs it. I sanity-check the outputs on Monday morning before the operating review.

This is the audit itself. Ten prompts, in order, with what to feed them and what the output should look like. You can copy this directly into Claude or ChatGPT with a project set up on your general ledger export.

Why weekly, not monthly

APQC’s cross-industry finance benchmarking puts the median monthly close at roughly 6 to 7 business days, with top-quartile teams under 5 (see CFO.com summary of APQC data). That means by the time your monthly close catches an error, the operating decisions built on the wrong numbers are already 25 days old.

Weekly beats monthly the same way rolling forecasts beat annual budgets. You can find related thinking in Rolling Forecasts vs Annual Budgets.

The audit runs on last week’s activity. It does not require close. It uses transaction-level data from the GL, the AP subledger, and the payroll register. That is it.

The setup (one time, 20 minutes)

Create a project in Claude or a Custom GPT with these context files loaded:

• Your chart of accounts (name + number + parent)
• Your last four quarters of trial balance
• Your top-50 vendor list with normal payment frequency
• Your employee roster with hire dates and comp bands
• Your revenue recognition policy (the actual policy doc)

Set the system prompt to something like: “You are a controller reviewing the general ledger of a $180M SaaS company. Flag anomalies. Do not summarize. Do not offer strategic commentary. Return only exceptions with the transaction ID, the anomaly, and the specific policy or historical pattern it violates.”

The 10 prompts

1. New vendor scan

Paste last week’s AP register.

“Compare the vendor list from these transactions to my top-50 vendor file. Return any vendor that is new (no prior payment history in the last four quarters) with an invoice over $10,000. For each, give me: vendor name, amount, GL account coded to, invoice date, and approver name.”

Why: New vendors over $10K are the highest-probability fraud and coding-error surface in AP. I have caught two fictitious vendors in six years using exactly this cut.

2. Duplicate payment check

Paste last 60 days of AP.

“Find any pairs of payments where vendor name is identical, amount is within 2%, and payment dates are within 30 days. Return the pair with a confidence flag: HIGH if amounts match exactly, MEDIUM if within 1%, LOW if 1-2%.”

Why: This is the $47,200 catch I mentioned. Vic.ai and other AP automation tools do this natively, but if you are still on QuickBooks or NetSuite without an AP overlay, the prompt is your net.

3. Round-dollar anomalies

Paste last week’s journal entries.

“Return every journal entry over $25,000 where the debit or credit is a perfectly round number (ends in ,000 or ,500). Flag those made by someone other than the controller and those posted after 6pm local time.”

Why: Round-dollar entries are the second-most-common signature of a manual override or plug. Legitimate entries almost never land on a round number unless they are accruals with a stated methodology.

4. Revenue recognition sanity

Paste last week’s new bookings, GL revenue accounts, and deferred revenue movement.

“For each new booking in the schedule, calculate what should hit revenue this week under ratable recognition (assume monthly straight-line unless noted). Compare to what actually posted. Flag deltas over 5% or $10,000.”

Why: If you are on ASC 606, the machine catches the reasonable subset of misapplications. It does not catch the hard judgment calls, but you knew that.

5. Aged AR triage

Paste your AR aging.

“List every invoice over 60 days past due where the balance exceeds $25,000. Group by customer. For each, tell me: original invoice date, days past due, last collection activity date if noted, and whether the customer has any outstanding disputes flagged.”

Why: Two hours of Friday phone calls on Monday’s list moves DSO more than any process redesign I have tried.

6. Accrual completeness

Paste last month’s PO log and last week’s invoices received.

“Return any open PO over $15,000 that has not received an invoice within its expected fulfillment window (assume 30 days unless stated). List these as accrual candidates for month-end.”

Why: The forgotten accrual is the sneakiest EBITDA hit. See How to Diagnose an EBITDA Miss for the downstream cost.

7. Payroll variance

Paste this week’s payroll register and last week’s.

“Compare gross wages, employer taxes, and benefit deductions week over week by department. Flag any variance over 3% at the department level or over $2,000 for individual comp changes. Return only the exceptions.”

Why: Payroll is 40 to 60% of most operating expense lines. A 3% variance nobody noticed is a real quarter.

8. Cash-to-book reconciliation

Paste bank statement activity and the GL cash account.

“Return any bank transaction over $10,000 that has no matching GL entry within 3 business days on either side. Also return any GL cash entry over $10,000 without a corresponding bank clear within 5 business days.”

Why: The unposted deposit or the timing-mismatch write-off is where the reconciliation-preparer errors live.

9. Journal entry review

Paste every journal entry over $50,000 from last week.

“For each entry, evaluate: is the supporting narrative in the memo field specific enough that a reviewer could reconstruct the source event without asking questions? Rate PASS, MARGINAL, or FAIL. For MARGINAL and FAIL, tell me what specifically is missing.”

Why: Bad memos are how future-you loses two hours in an audit. Fix now.

10. The manager read

This one is for you, not the machine.

“Summarize the nine outputs above in exactly 200 words for a CFO who has five minutes. Lead with the dollar amount at risk if unresolved. Do not summarize what you did. Summarize what I need to decide.”

Why: If your controller is running the audit and you are consuming it, this closes the loop.

What the output looks like in practice

On a normal Friday, seven of ten prompts return “no exceptions.” The other three return between one and four items each. Total items to review Monday morning: usually five to twelve. Total time on my end: 30 minutes.

The week I found the $310,000 stock comp miss, the payroll variance prompt flagged a $12,400 delta in the executive department that traced back to the missing accrual. The prompt did not calculate the accrual. It told me to look at that department.

The counter-argument

The reasonable objection: “This is just a checklist. The controller should already be doing this.”

Yes. The controller should be doing this. Most are not, or are doing three of the ten and missing the other seven. Deloitte’s controllership research over the past several years has documented a steady shift of controller time away from transactional review toward systems work, exception handling, and reporting mechanics as automation matures. The audit did not go away. It got quieter.

The second objection: “I do not trust the LLM to catch what a human catches.” Correct. The LLM does not replace the human. It surfaces the exception list. The human still decides. If your controller is spending two hours a week reviewing the output instead of 20 hours a week hunting for anomalies, that is the trade. The hit rate on false positives runs about 15 to 20% in my experience, and every one of those is worth the two seconds it takes to dismiss.

The third objection: “What about data privacy?” Legitimate. Keep the vendor names, keep the amounts, strip employee-level payroll to department aggregates before pasting. See The CFO’s Guide to Uploading Financials to LLMs Without Privacy Risk.

What this connects to

The 2026 AI CFO Benchmark found that 41% of 10-K filings mention AI in 2026 YTD. Most of that mention is marketing. The controllers actually using it in a repeatable weekly cadence are a much smaller number. The compounding is real. Twelve months of running this audit puts your team in a place where the anomaly detection is faster than any tool you would buy for it.

If you want the full set of prompts formatted as a project ready to paste, plus the 25 other prompts I run for board work, forecasting, and diligence, the AI-Native CFO Prompt Pack Pro has all of them.

Note: Company details in this piece have been anonymized. Any figures drawn from Spencer’s advisory work with middle-market and PE-backed finance teams are directional.