insomnia.club back to site
[ playbook · finance and accounting ]

An AI expense audit agent that reviews every report, not a sample

Most companies audit a sample of expense reports, if any, because reading receipts against policy is tedious. An agent can review every line, compare it to the receipt and the written policy, and give finance a short list of items that deserve a question.

who owns it

Controller or finance manager, with department heads approving their teams' spend

what starts it

An employee submits an expense report or a corporate card transaction settles

01the problem and who owns it

Managers approve expense reports in a few seconds between meetings. Finance spot checks some of them at month end. Between those two glances, duplicate submissions, receipts that do not match the claimed amount, alcohol charged to a client meal with no client, and personal purchases on corporate cards slip through, mostly by accident and occasionally not.

The controller owns policy enforcement but cannot staff a full review. Employees, meanwhile, find the policy vague and ask the same questions repeatedly, so some violations are honest confusion that better feedback would fix.

02what the AI does, step by step

  1. Read the line and the receiptFor each expense, the agent reads the receipt image and compares merchant, date, amount, currency, and itemization to what the employee entered. Mismatches, missing receipts above the policy threshold, and edited-looking images are noted.
  2. Check against written policyYour expense policy, in plain language, is the reference. The agent checks per diem and meal limits, lodging caps by city if you set them, required attendees for client entertainment, and categories that are never reimbursable, citing the policy clause for each finding.
  3. Look across reportsIt searches across employees and periods for duplicates: the same receipt submitted twice, a card charge also claimed as out-of-pocket, or two colleagues expensing one dinner. It also notes spend that is unusual for that role or trip.
  4. Grade each findingFindings are graded as informational, needs explanation, or likely violation. Small, clearly honest issues can trigger a friendly note to the employee explaining the rule instead of a finance escalation.
  5. Ask the employee firstFor needs-explanation items, the employee gets a specific question, such as who attended this dinner, and their answer is attached to the report before a person reviews it.
  6. Send the queue to financeFinance reviews likely violations and unresolved questions in one queue, with the receipt, the policy citation, and the employee's response side by side, and decides the outcome.

03systems it connects to

04human checkpoints

05what to measure

06risks and guardrails

07build vs buy

Modern expense and card platforms such as Ramp, Brex, Navan, and Concur include policy rules and some AI-based receipt checks. If your policy is simple and your spend sits inside one of those platforms, turn on what they offer first.

A custom agent makes sense when the policy has nuance those rule builders cannot express, such as client billable rules, project codes, or union travel provisions, or when you need checks across several systems, for example expenses against the CRM to confirm a client meeting happened.

Browse every finance and accounting playbook or the full library.

want this running in your business?

Share your expense policy and a recent month of anonymized reports, and we will show you the findings an agent would raise before you decide whether to put it into production.

See how we deliver it: ai agent development.

book a call drop your number

info@insomnia.club