How to review employee expense claims with an AI agent

Review every claim, not a sample. The agent reads each receipt, compares it with the claim, applies each policy rule, looks for duplicates across claimants and periods, and writes a short finding: clean, query or exception. Line managers approve the business purpose, finance decides exceptions, and suspected misconduct goes to a named person in finance or HR.

What does expense review involve, and who owns it?

Employees claim reimbursement for business costs: travel, meals, accommodation, mileage, subscriptions. Policy sets what is allowed, at what limits and with what evidence. Review checks each claim against that policy and the receipt. Most finance teams can only sample claims, so most are never read by anyone in finance.

The line manager approves that the cost was for the business. The finance or expenses team checks policy compliance and evidence. The policy owner, usually the financial controller, sets the rules. HR becomes involved when a claim looks dishonest. The agent's contribution is coverage: it reads every claim with the same attention. Within the finance bounds set out in AI agents for finance, that is the argument made in expense agents enforce policy consistently; people handle the exceptions.

What does the agent prepare?

1. Agent, depth 1: reads the claim and receipts. Merchant, date, amount, currency, items, tax, and the claimant's stated purpose.
2. Agent, depth 1: compares receipt to claim. Amount, date and merchant agree; the receipt is itemised where policy requires; the receipt is not a card slip alone.
3. Agent, depth 1: applies policy rules. Category limits, per diem rates, mileage rates and distance plausibility, class of travel, alcohol, attendees for meals, advance approval for certain spend.
4. Agent, depth 1: checks duplicates. The same receipt claimed twice, by the same person or by colleagues for a shared meal; a claim also paid on a corporate card; the same mileage claimed by two people in one car.
5. Agent, depth 2: writes the finding. Clean, query (a specific question for the claimant) or exception (a policy breach or anomaly), with the policy clause and evidence.
6. Agent, depth 2 or 3: sends queries. Standard questions to the claimant, such as "please add the attendees for this meal", once certified.
7. Person, line manager: approves business purpose. Sees the agent's finding with the claim.
8. Person, finance reviewer: decides exceptions. Approve with a note, part-approve, or reject.
9. Person, finance lead or HR: handles suspected misconduct.

What does the named person approve?

The line manager approves that the expense was incurred for the business, which only someone who knows the work can judge. The finance reviewer decides exceptions: whether a breach is reasonable in context, such as a late-night taxi after a delayed flight, or should be recovered. The agent informs both decisions and makes neither. Where you want the agent to approve the simplest claims itself, letting an AI agent approve low-value expense claims sets out how; whether it should ever reject claims is covered in can an AI agent approve or reject expense claims.

Evidence for each claim: the finding, the policy clause and version, the receipt check result, the duplicate check result, the manager's approval and the reviewer's decision.

What does a finding look like?

Take an illustrative claim for a team dinner of 480 with four attendees. The agent finds the receipt matches, the per-head cost is within the policy limit, and alcohol is itemised at 120, which the policy allows only with a director's approval. It also finds that one attendee claimed 95 for a meal at the same restaurant on the same evening. Finding: exception, with the alcohol clause and the possible duplicate shown side by side. The reviewer confirms the second claim was a separate client dinner earlier that evening and asks for director approval on the alcohol.

Where does it usually go wrong?

The agent becomes the accuser. A finding is sent to the claimant as "you broke policy". Findings go to reviewers; only neutral, standard questions go to claimants.

Policy says "reasonable". The agent cannot apply an undefined limit. It should flag the gap to the policy owner rather than invent a number.

Shared costs. One person pays for a group and others claim parts of it too. Checking duplicates across claimants is the agent's main advantage over sampling.

Receipt images trusted. Edited or reused receipt images exist. The agent compares with card data where available and flags receipts reused across claims.

Tool access too wide. The agent's account can change approval chains or policy settings. Scope it to read and comment, as in connecting an AI agent to your expense management tool.

What should you measure?

• Claims with exceptions, by rule.
• Duplicates found, by type.
• Queries answered and how quickly.
• Amounts recovered after review.
• Findings overturned by reviewers, which show where the agent is reading policy wrongly.

Where Pryme fits

In Pryme, reading receipts, applying policy and sending standard queries are separate capabilities, each certified through Attest on your own past claims. Card feeds arrive signed by the issuer and receipt images as external, so a finding shows whether it rests on the card provider's data or on an uploaded picture. Every finding and decision is filed.