The Expense Report Nobody Fills Out, Audited 100% of the Time

Two things are true about travel and expense. Employees hate filing reports, and finance hates auditing them. AI happens to be good at both problems at once, which is rarer than it sounds.

Start with the employee side, because that's where adoption lives or dies. Snap a photo of a receipt and IDP pulls the merchant, date, amount, currency, and tax — then auto-fills the expense line and often guesses the category from the merchant. Corporate card feeds get matched to the right receipt automatically. Done well, filing an expense report goes from a dreaded twenty-minute chore to a few taps. The submission itself becomes almost incidental.

The bigger prize is on the audit side

Here's the shift most people miss. Traditional T&E audit samples — you check maybe 10% of reports and hope the other 90% behave. AI flips that. A GenAI-assisted policy engine can read every line item against your actual policy and flag violations across 100% of submissions: the meal over per-diem, the flight booked outside the advance window, the suspiciously round "cash tip," the duplicate receipt submitted twice three weeks apart.

Full-population auditing changes the controls conversation entirely. You're no longer extrapolating from a sample; you're seeing everything. And because it runs pre-reimbursement, you catch violations before the money's gone rather than clawing it back later — which anyone who's tried to recover an over-payment from a departed employee will tell you is worth a lot.

The numbers that move: manual data-entry time per report drops toward zero, approval cycle time shrinks because clean reports flow straight through, and out-of-policy spend becomes visible instead of buried. We often find the flagged spend was never malicious — it was just never caught.

What to watch for

Receipt OCR still stumbles on crumpled thermal paper, foreign-language receipts, and photos taken in a dim restaurant. Build a graceful fallback so the employee can correct a misread field without rage-quitting the app.

Policy language is where GenAI needs a firm hand. Real policies have nuance — "reasonable" client entertainment, exceptions for certain roles, regional per-diems. If you feed the model vague rules, you'll get vague flags and a lot of false positives that erode trust. We translate policy into specific, testable checks and tune the thresholds before turning on enforcement.

And a human still owns the judgment calls. The AI flags; a person decides whether a flagged item is a genuine violation or a legitimate exception. Auto-rejecting on the model's say-so is how you turn a controls win into an employee-relations problem.

One quiet benefit worth naming: when people know every report is audited, behavior changes on its own. The deterrent effect does as much work as the detection. Get the employee experience right and the controls come almost for free — that's the combination that makes T&E one of the easier AI wins to justify.