AI-generated receipts are here. What changed for expense review

For most of the history of expense review, a fake receipt was a craft project. Someone downloaded a template, typed in a merchant and an amount, and printed or exported the result. Templates had tells: the same layout for two unrelated restaurants, a font no point-of-sale printer uses, columns that did not quite line up. A reviewer with a good eye and a few minutes could catch a fair share of them.

That changed in the spring of 2025. Leading AI models gained the ability to render legible, correctly spelled text inside images, and within days people were posting restaurant receipts the models had produced from a one-line prompt, some crumpled or stained to look handled. The same models also edit: give one a photo of a real receipt, ask for a different total or date, and most of what comes back is genuine. The result looks like a photo of a real receipt, because to a large degree it is.

What the numbers show

The shift is now visible in audit data. An expense-audit software provider, AppZen, reported that receipts made with AI image generators were about 14% of the fake receipts it detected in September 2025, up from none a year earlier. By mid-May 2026 the same company put the share at 70.8%: 1,471 AI-generated receipts submitted by 745 employees at 174 companies, against 29% for the template fakes that had been 95% to 100% of its catches a year before.

Two details in that data matter for reviewers. First, the AI-generated fakes are small. They averaged about $100, with a median around $32, against about $182 for template fakes. Amounts like that sit comfortably under most auto-approval thresholds and rarely draw a second look. Second, the figure describes the make-up of fraud that was caught, not how much fraud there is. A receipt that nobody flags does not appear in anyone's statistics.

Survey data points the same way. In a survey of 2,000 workers in the United States and the United Kingdom published by the expense software company Emburse in June 2026, 40% of U.S. respondents and 29% of U.K. respondents said they had used AI to create a fake business receipt. The reasons ranged from inventing a purchase that never happened, to inflating a real one, to recreating a receipt that had been lost.

Why the old visual checks stop working

A generated or AI-edited receipt fails differently from a template. The typical template tells disappear: the layout matches what that kind of merchant prints, the fonts look like printer fonts, the paper curls and casts a shadow. A reviewer who asks “does this look like a real receipt?” will get the answer “yes” far more often than before.

What generated receipts still get wrong tends to be content, not looks:

• Arithmetic. Line items that do not add up to the subtotal, tax that does not match the local rate, a total that ignores the discount printed above it.
• Merchant details. An address that does not exist, a tax or registration number in the wrong format, a phone number with the wrong area code for the city.
• Internal consistency. A table number on a takeout order, a card ending that differs from the card used, a time stamp outside opening hours.
• Context. A meal in a city the employee was not in that day, or no matching card transaction at all.

When a fake starts from a real receipt rather than from nothing, the merchant, the layout and most of the line items are genuine. What changes is the amount, the date or the number of covers. That is where the review has to look.

Why it matters more in life sciences

In pharmaceutical and medical device companies, a receipt is often more than proof of a cost. A meal receipt or a photo from a meeting with healthcare professionals can be the main record that an interaction happened as reported: who attended, where, and what was spent per person. In the United States those records feed the Open Payments program, under which manufacturers report meals and other transfers of value to physicians and teaching hospitals. In other markets they feed local transparency rules and internal spend limits.

Regulators have also made clear that the events themselves draw scrutiny. The HHS Office of Inspector General's November 2020 Special Fraud Alert on speaker programs listed, among other warning signs, meals beyond modest value and attendees with no legitimate business reason to be there. A fabricated or altered receipt in that setting is not a small expense problem. It can hide a transfer of value that should have been reported, or an event that did not take place as described.

Where to focus review effort

No compliance team can forensically examine every receipt. The data above suggests where the risk concentrates:

• Claims with no matching card transaction, such as cash payments and “lost card statement” claims.
• Amounts just below an auto-approval or per-person limit, especially when they recur.
• Replacement receipts submitted after the original was reported lost.
• Meetings with healthcare professionals where a photo or a single receipt is the only evidence of attendance.
• Receipts submitted as screenshots or forwarded copies when the employee should have the original.

What a modern review looks like

The answer is not a single tool. It is a review in layers, each of which catches something the others miss.

1. Match the claim to independent data. A genuine purchase leaves a card transaction with its own date, merchant and amount. Matching against that record catches fabrications regardless of how good the image looks.
2. Get the best file you can. Every forward, screenshot and scan erases some of the traces that detection depends on. Ask for the original before you analyze.
3. Examine the image. VeriSight runs three checks on one upload: an AI check that scores the image for AI generation or AI editing, a tamper check that outlines regions edited with image software, and a visual review with four views of the image that make typed-over text and pasted regions easier to see. Our guides on reading the AI score and spotting typed-over fields cover how to read each one.
4. Check the content. Do the sums, check the merchant, compare with the employee's itinerary and with other claims from the same trip.
5. Document what you found and how. Record which file you analyzed, where it came from, what each check returned and what you corroborated it with. A finding that names its evidence holds up in a conversation with the employee; a hunch does not.

Image generators will keep improving, and so will detection. What does not change is the principle: a receipt is a claim about a transaction, and the strongest review tests that claim against evidence the claimant did not produce.