Expense audit software that reads every receipt, shows why one is doubted and waits for you to decide
Every report checked, not a sample, and no forged receipt paid without proof of payment.
See one case, screen by screen ↓Expense Report Audit is expense audit software that checks every expense report and receipt before the pay run, not a 5–10% sample. Seven AI agents read each receipt and check it for signs of AI generation, against the policy and city rates, against the corporate card feed and across other people's claims. A cross-checker weighs those findings and proposes a decision. The expense auditor decides every doubted line and signs it with reason codes.
How to choose AI solutions for back-office and enterprise work →
Updated
Most expense teams audit a random 5–10% sample of reports and let rules check the rest. The rules check the amount the employee typed, not the receipt image. Free image tools now produce a convincing restaurant receipt in seconds. A forged dinner that stays inside the meal rate and the entertainment limit passes every rule, and the line manager approves it on a phone in under a minute. It is paid in the next run before anyone has looked at it.
When a report is pulled for audit, the auditor works across the expense system, the card portal and the policy PDF. A queried employee often reads the email as an accusation, and nobody tracks when the reply is due. Card spend with no receipt gets chased monthly, after the 120-day return window has gone. A dinner split between two colleagues, or one hotel room claimed by two people, can't be seen one report at a time. When internal audit later asks who decided and why, the evidence is spread across shared drives.
Estimated hours for one expense audit team in one weekly pay cycle. Today's hours cover a sample; with the solution, the hours cover every report.
Seven agents read, inspect and cross-check every receipt against the card feed, the policy and each other; the expense auditor decides and signs.
Kestrel & Vane Group pays expenses in a weekly run with a Thursday 12:00 cut-off. On Wednesday 14 October, Elena Varga, expense auditor for the Americas, opens the solution. The agents have already read every report sent since Thursday. This is how she works through the morning with them, screen by screen, in the working solution.
Her day starts with a briefing, not a dashboard. Since Thursday the agents have read 1,284 reports and 8,200 receipts. 1,171 cleared on their own, 38 need her before the cut-off, and $18.7k is held until she decides. Each agent writes a short note with a proposal and one-click answers, and the cross-checker's note comes first: Marcus Delaney's Chicago trip, ER-20931, $2,158.52.
“I propose paying the 11 on Thursday and asking Marcus about the dinner. The message is drafted for you.”
The report opens with its receipts laid out side by side. Each carries small badges for the forensic likelihood, the card match and the policy check. Eleven say Cleared. R6, a $312.40 dinner at Brannock's Bar & Chophouse, has 94%, a crossed-out Card badge and a Font badge, and needs a decision. Beside the receipts the agents show what they found, the evidence and what she is asked to decide.
Four numbered marks on the receipt show why it is doubted. The digits of “312.40” are in a different typeface, 11% wider than every other figure. $268.00 plus $31.49 tax is $299.49, which leaves $12.91 unexplained, and there is no tip line. The paper texture is generated, with no thermal-printer dot pattern. The file is a PNG with no camera data, made on Tue 13 Oct at 17:58, a week after the dinner.
“Escalate: the forensic flag is not cleared by any payment record (rule X-1).”
The breakfast receipt R5, Canal Street Bagel Co. at $18.75, also scored 41% on forensics because the photo was taken in low light. The card feed shows TX-88421006, with the same merchant, the same amount and the same day, so rule X-2 cleared it with a note and no person had to look. R6 has nothing like that: the claim says it was paid on a personal card, no statement is attached, and the card shows no spend after 18:00 that day.
“Cleared: forensic doubt answered by an exact card match (rule X-2).”
The Brannock's receipt has the same layout as two others: Ana Ruiz's $96.40 at Bayou Ember Grill in Houston and Kofi Mensah's $91.85 at The Gilded Ferret in New York. All three have the same header block and a typeface change in the total (+11%, +10%, +12%), and none has a payment record. The pattern finder files them as pattern PT-11, $500.65, with 92% confidence, and suggests sending all three to internal audit.
“No one is named to a manager until you confirm.”
Elena accepts the recommended option: approve 11 lines for $1,846.12 and send R6 back to Marcus. The reason codes are already ticked: RT-01, receipt could not be verified as genuine, and RT-02, no proof of payment. The explanation agent drafted the message in plain words, and she can edit it before signing. Rejecting R6 as a confirmed generated receipt would need Rachel Lindqvist, the T&E manager, to countersign.
The report moves to Waiting on the employee and R6 shows “Sent back · EX-3401”. In the conversation beside it, the explanation agent says it sent EX-3401 in plain words and will remind Marcus on day 3. The pattern finder notes that this is Marcus's second returned line in 180 days, so his line manager now has sign-off SG-5524.
“ER-20931 signed · 11 paid · 1 sent back”
Marcus sees that 11 of 12 receipts are approved, $1,846.12 paid with his salary on Thu 15 Oct, and that one needs his help by Wed 21 Oct. The message says the receipt could not be confirmed and explains why in plain words. He has four ways to answer: add the statement line, add the emailed receipt, remove the line, or say something is wrong.
“A person reads every reply and makes the decision; nothing is decided by the system alone.”
SG-5524 shows Graham both cases side by side. ER-19877 on 18 Aug was a lunch receipt re-used with a new date, and Marcus removed it. ER-20931 is the dinner that could not be verified. Graham records a decision: he has talked to the employee, he is asking HR for advice, or no action, with a reason. Graham raised neither case, and he cannot sign what he raised.
“Nothing here says Marcus did wrong on purpose; the auditor decided only that the receipt could not be verified.”
The audit trail lists every step on ER-20931: the report reader opening it, the four checks, the cross-checker's escalation, Elena opening R5 and R6, her signed decision with RT-01 and RT-02, the message to Marcus and the payroll queue for $1,846.12. Each entry carries a fingerprint chained to the one before. Verify the chain reports “Chain intact: 21 of 21 entries”, and an evidence pack can be exported for the report.
The same forged dinner, seen by the five people it reaches: how their week used to go, and how it goes now.
Seven agents read, inspect and cross-check every receipt against the card feed, the policy and each other; the expense auditor decides and signs.
Reads every report and every receipt: merchant, date, items, tax, tip, total and currency, in any language, and ties each receipt to its report line.
Inspects the image itself for generated textures, missing printer patterns, typeface changes, edited digits, sums that don't add up and the file's data, and checks it against earlier receipt images.
Checks every line against policy TE-4 and the city rates: receipts, payment, meals, entertainment, alcohol, lodging, mileage, timing and weekends. It reads a new policy version into proposed rule changes.
Matches each line to the corporate card feed. It finds claims with no payment record, card spend with no receipt, amounts that differ and lines paid twice, and asks for missing receipts.
Looks across people and 12 months of receipts for the same receipt claimed twice, bills split to stay under a limit, shared fake layouts, round-number habits and repeat cases.
Weighs the four checks against each other with the cross-check rules. A forensic doubt with no payment record goes to the auditor, and one answered by an exact card match is cleared with a note. It proposes a decision with its confidence.
Writes the request to the employee in plain words and chases it on the service level. It reads the reply and any statement image, and says whether the explanation holds.
Decides with reason codes; the T&E manager countersigns rejections. The agents propose; a named person decides.
Every screen has a box for asking the agents about a report, a receipt, a person or a rule, or telling them what to chase. Answers cite their sources and offer the next step. Decisions stay with an auditor.
What should I do first?
Marcus Delaney's Chicago trip, ER-20931. It is the only report with a likely generated receipt and no payment record, and it must be decided before the pay-run cut-off, Thu 15 Oct 12:00. It takes about two minutes: one receipt to look at, one decision to sign.
How much is held from Thursday's pay run?
$18,719.85 on 71 reports. 38 need an auditor, 27 wait on the employee and 6 are with line managers. The largest is Daniel Kurz's June trade fair ($2,240.55, sent after 60 days). Everything else is paid on Thu 15 Oct.
Why did the breakfast receipt clear?
The card answered the doubt. The photo of R5 is dark, so forensics gave it 41%. But card ending 4471 shows Canal Street Bagel Co., $18.75, the same morning: rule X-2 clears it, with a note.
Are AI-generated receipts getting worse?
Yes: from none a week in July to 13 in the week of 5 Oct, 54 in 12 weeks, about $115 each. Every one passed the policy rules; the image and the missing payment gave them away. Three share one layout across three cities.
13 screens from the working solution, on its sample data. Pick one to see it large.
The rules are read from your own policy, and the thresholds are yours to move. Before you send a change, you can see what it would have done in the last 90 days. Changes go to the T&E manager before they take effect.
| Setting | Default | Choose from |
|---|---|---|
| Escalate a receipt as likely generated or altered | 60% forensic likelihood | At or above it, a receipt needs a payment record to clear (rule X-1) |
| Escalate meals and entertainment | 60% | 55% proposed, waiting for the T&E manager (SG-5517) |
| Keep a note, clear on a card match | From 30% | Between this and the escalation line, an exact card match clears the receipt (rule X-2) |
| Auto-clear small policy breaches | Under $25 with a card match | Cleared with a note to the employee (rule X-4) |
| Quality sample of cleared reports | 2% | Re-checked by a person to measure the agents |
| When a policy rule is hit | Set rule by rule, e.g. R-01 Return, R-03 Explain, R-07 Auditor | Explain · Return · Auditor; integrity rules always reach a person (rule X-3) |
| Explanation requests | 5 working days | Reminder on day 3; line manager copied on day 5 |
| Repeat cases | Second returned line in 180 days: line manager | Third: the T&E manager and HR |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample expense data · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published figures (5–10%: the default and example random audit sample rates in the documentation of widely used expense tools; 70.8%: one expense-audit vendor's own data on the fake receipts it flagged, reported by the Ohio Society of CPAs, June 2026; 40%: a 2026 survey of 2,000 US and UK workers by an expense software company, US respondents). Accountable-plan windows of 60 and 120 days: IRS Publication 463. Kestrel & Vane Group, its merchants and the people named on this page are fictional: characters and sample data in the working solution.
Expense Report Audit is expense audit software that checks every expense report and receipt before the pay run, not a 5–10% sample. Seven AI agents read each receipt and check it for signs of AI generation, against the policy and city rates, against the corporate card feed and across other people's claims. A cross-checker weighs those findings and proposes a decision. The expense auditor decides every doubted line and signs it with reason codes.
The receipt forensics agent looks at the image itself: generated paper texture, a missing thermal-printer pattern, a total set in a different typeface, sums that don't add up, and file data such as no camera details or a creation date after the expense. It marks each finding on the magnified receipt. A forensic doubt is never a verdict on its own: it needs a card, bank or merchant record before the line can clear.
A doubt that the card feed answers is cleared with a note and doesn't reach a person. One example is a dark breakfast photo that scored 41% but matched the card exactly. In the working solution the agents cleared 1,171 of 1,284 reports on their own, auditors cleared 9.6% of what the agents flagged, and a 2% quality sample showed 99.0% agreement.
Yes. The pattern finder compares each new receipt with 12 months of receipts across all employees. It finds one dinner split into two claims to stay under a limit, one hotel room claimed by two people, a receipt re-used with a new date, and the same fake layout used in different cities. Linked receipts are shown side by side.
The explanation agent drafts the request in plain words, and it is sent only when the auditor signs. The rest of the report is paid on time. The employee can add the statement line, add the emailed receipt, remove the line or say something is wrong. The agent chases on a five-working-day service level, reads the reply and says whether it holds.
Yes. The policy checker reads your policy into rules you confirm and checks every line against them: itemised receipts, corporate card use, city meal rates, entertainment and alcohol limits, lodging caps, mileage, the 60-day submission window and weekend spend. City rates are maintained in settings. A new policy version produces proposed rule changes for the T&E manager to approve.
Yes. The agents never approve or reject a doubted line, never accuse anyone and never send a request without a signature. The expense auditor decides with reason codes. The T&E manager countersigns rejections as confirmed forgery, recoveries, policy exceptions and setting changes. Nobody signs what they raised, and every step is on a chained audit trail.
The Agentic Solution Engine builds and deploys it from your requirements and documents, including your T&E policy, your expense system and card feed, and a month of past reports. It goes live once every quality gate has passed. First-time setup includes a trial on the month before go-live, so you see what the agents would have found. In the working solution that was 5,384 reports, 471 flagged and 22 likely generated receipts.
Blocked invoices released before late fees and supplier holds, every fix checked twice.
1-min demoVendor Payment Fraud PreventionEvery bank change proven by a call-back before a single dollar moves.
1-min demoDuplicate Payment Detection & Recovery AuditTwins stopped before the payment run, paid ones recovered in house, every claim signed.
1-min demoWe'll run Expense Report Audit on a month of your own expense reports, receipts and card feed.