AI Expense Report Review: Flag Exceptions, Leave Approval to People
How an AI agent checks expense claims against policy and flags missing receipts for a manager, without approving or paying any claim on its own.

An AI agent can review expense reports before a manager opens them by checking each claim against written policy, matching it to its receipt, and marking anything that needs a human decision. The agent returns a short review packet with a reason and evidence for every flag. The manager approves, rejects or asks for more information, and payment stays with the finance process that already owns it.
Key takeaways
- The agent compares each claim with written policy and attaches the evidence for every flag it raises.
- Missing receipts become a defined request to the employee, never a value the agent fills in by guessing.
- Approval, rejection and payment stay with named people and the existing finance system.
- Each flag carries a reason, an owner and a deadline so no exception waits unseen.
- Begin in observation mode, compare the agent's flags with manager decisions, then widen the scope slowly.
Why expense review suits an agent
Expense review repeats the same questions on every claim: is there a receipt, does the amount match, is the category allowed, do the dates fall inside the trip, has this line already been submitted. Most companies have written answers to these questions, which makes them testable. A manager still needs to judge the cases the policy does not settle, such as a client dinner that ran long or a seat upgrade on a medical trip. The useful split is that the agent does the checking and the manager does the judging.
The task is also narrow enough to measure. The agent either raised a real exception with clear evidence or it did not, and the manager's decisions show which. That kind of observable success criterion is what a first engagement should define. Where the packet goes once it is ready is a routing question, covered in AI Approval Workflow: Route Each Request to the Right Approver.
Define the policy rules the agent is allowed to test
Before an agent reads a claim, the policy needs to be written as rules someone could check by hand. "Meals should be reasonable" cannot be tested. "Meal claims above the limit set by finance need an itemized receipt and a named attendee list" can be. A workable rule set usually covers:
- Receipt presence, and whether the receipt amount matches the claimed amount.
- Category limits, with each limit value stored in the policy document that finance maintains.
- Travel dates checked against the approved dates in the trip record.
- Duplicate detection by merchant, date and amount across the same employee and period.
- Categories that always need a manager review, such as gifts, alcohol or anything outside an approved list.
Rules that depend on context the agent cannot see should return unknown and go to a person. If a meal is a business expense depends on attendees, and the attendee list is missing, the flag should say exactly that. The agent should never fill the gap with its best guess. Give each rule a version number so the review log shows which version was applied to each claim.
Build the review packet step by step
The agent's output is a packet for the manager, not a decision. A sequence that works in practice looks like this:
- Intake: read each submitted claim with the employee, date, category, amount and currency as recorded by the expense tool.
- Match: link each line to its receipt and record whether the amounts agree.
- Check: run each policy rule and record pass, flag or unknown for each line.
- Compare: look for duplicates and for lines that repeat an earlier claim.
- Draft: write one line per flag with the rule, the evidence and the decision the manager needs to make.
- Route: send the packet to the approver defined in the approval chain, never to a recipient the agent chose.
- Log: store what was checked, what was flagged, which policy version was used and when the packet was sent.
Each flag should be readable on its own. "Hotel night outside the approved travel dates in the trip record, with the trip dates shown beside it" gives the manager something to act on. "Date issue" does not. Packets that sit without a decision need the ownership and deadline discipline described in Exception Queues for AI Workflows: Owner, Reason, Deadline.
Handle missing receipts without guessing
Missing receipts are the most common flag and the easiest to mishandle. A tempting shortcut is to match a claim to a card transaction and treat that match as proof. A card statement shows that money left an account. It does not show what was bought, who attended, or whether the purchase was for company business. The agent can note a matching transaction as context for the manager, but it should not mark the receipt as found.
When a receipt is unreadable, the agent marks the field as unreadable rather than reconstructing the merchant name or tax from partial text. The request to the employee names the exact missing item, with the date and amount as recorded, and sets a deadline. If the receipt does not arrive in time, the packet goes to the manager with the claim marked receipt missing, and the claim stays open.
Receipts are also an input risk. A scanned document can contain text that tries to instruct the system, such as a note asking for everything to be marked approved. The agent should treat document contents as data to be checked, never as instructions. The [OWASP LLM Prompt Injection Prevention Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) explains why untrusted input needs that separation.
Keep approval and payment with people
The safest permission design gives the agent read access to claims and receipts, write access only to flags, drafts and logs, and no access to approve, edit amounts or trigger payment. This is the least privilege principle that the [OWASP Authorization Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) describes: each actor gets only the access its job requires. A short decision guide helps the team agree on the boundary before the agent starts.
- The agent may read claims, attach receipts, set flag reasons and draft requests to employees.
- The manager decides approval, rejection, partial approval and any policy exception.
- Finance decides payment timing, ledger coding and tax treatment questions.
- The agent never approves a claim, edits an amount, marks a receipt as verified or reviews a claim it is connected to.
- Any claim where the reviewer has a conflict, such as their own expenses, routes to a second approver.
The separation matters for audit too. When someone asks why a claim was paid, the answer should name a person and a decision, not a tool. Teams that run invoices through the same approvers often find the same boundary useful, as set out in AI Invoice Intake With Human Approval: A Workflow Guide.
Illustrative scenario
Illustrative scenario, not a client case: an employee submits a monthly report with four lines. One taxi receipt is missing. One meal exceeds the limit and has an itemized receipt but no attendee list. Two hotel nights fall outside the travel dates in the trip record. A restaurant line matches a line on an earlier report.
The agent produces one packet with four flags. The taxi line gets a receipt request with a deadline. The meal gets the policy reason and a note that the attendee list is needed to decide whether it is a business meal. The hotel nights are shown beside the trip dates they fall outside. The duplicate restaurant line appears with both claim references. The manager approves the taxi once the receipt arrives, asks for the attendee list before deciding the meal, rejects the hotel nights as personal travel and removes the duplicate. The agent's flags and the manager's decisions then become the feedback used in the next evaluation.
Measure the review and how Be Human runs it
Measure the review by what the manager does with it. Useful signals include how often the manager agrees with each flag type, how many exceptions a periodic audit sample finds that the agent missed, how long packets wait for a decision, and how many employee requests pass their deadline. A flag type that managers always dismiss needs a rule change, and a type of exception that slips through needs a new check. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) treats this kind of ongoing measurement as part of managing AI risk. Running the agent in shadow mode first, where it produces packets but changes nothing, gives the team this evidence before any wider role is granted, as described in Shadow Mode for AI Agents: Watch Before You Grant Autonomy.
In a Be Human engagement, the first step is a written role for the expense reviewer: the policy rules it may test, the systems it can read, the actions it cannot take and the success criterion the manager agrees to. Managed operations then cover monitoring of flags, recurring evaluations against manager decisions, permission reviews when expense tools change, and controlled updates to rules when finance revises the policy. The aim is an agent that prepares decisions well and leaves those decisions to the people who own them.
If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
How does an AI agent review an expense report? It reads each claim, matches lines to receipts, checks them against written policy rules and drafts a packet listing each flag with its evidence. The packet goes to the approver defined in the company's approval chain. The agent does not decide the claim.
Can an AI agent approve small claims automatically? Not in a first engagement. Automatic approval is a permission change that should follow observation data, a documented policy and a named owner for any mistakes. Even then, the approval would rest on a rule the company wrote, not on a judgment the agent made.
What happens when a receipt is missing? The agent drafts a request to the employee that names the exact missing item and sets a deadline. If the receipt does not arrive, the packet goes to the manager with the claim marked receipt missing. The agent does not reconstruct the receipt from a card statement.
Which parts of expense review should stay fully manual? Judgment about business purpose, exceptions the policy does not cover, disputes with employees, tax treatment and any suspicion of fraud should stay with a named person. The agent can still gather the context and route the case with its evidence attached, so the person starts with the facts rather than a search.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


