Forging a convincing receipt used to take real skill. It doesn’t anymore.
Anyone with a phone can now use AI to generate a receipt for a restaurant that exists, on a date that makes sense, with a total just under an approval threshold. The best AI-fake receipts include details people associate with authenticity, like a coffee stain or a crease down the middle of the paper. The visual “tells” that auditors were trained to spot, like misaligned logos or a mismatched font, are now easy for an AI model to get right.
In early 2025, no expense automation tools were regularly flagging receipts with AI-generated images. By mid-May 2026, AppZen forensics indicated AI-generated receipts accounted for 70.8% of fraudulent receipts flagged by our platform. In the one year from May 2025 to May 2026, our models detected 1,471 AI-generated receipts. Together, those receipts represented $148,143 in expense claims and came from 745 employees across 174 companies.
Although the individual amounts were small, averaging $101 and with a median of about $32, what mattered was the pattern we were seeing and the change in the methods being used.
The question finance leaders often ask me is a fair one. If the receipt’s picture is perfect, what is left for an AI-powered system to detect?
More than you might expect.
What AI Finds in the Receipt Image
The first examination is forensic.
Generated images can carry a history. Metadata fields may indicate the tool that was used to produce them. A photograph of a piece of paper has a compression history, a light source and a focal plane. A rendered image may lack some of those characteristics, and simulating all of them convincingly is difficult.
The rendering itself can leave marks. Image models remain imperfect at generating dense, small type. Repeated characters may drift, and glyph shapes can vary where a thermal printer would print them identically. Visual “noise” intended to mimic faded print can sit on top of the text rather than appearing as part of the print itself.
None of this proves anything alone. Together, however, these signals can turn an unremarkable receipt into one worth a closer look.
What AI finds in the Transaction
The second layer of examination has little to do with how the receipt looks.
Real receipts are outputs of a real system, so they obey arithmetic. Line items add up to a subtotal, the tax matches the rate in that jurisdiction, and a tip is a plausible percentage of the right base. Fabricated receipts can fail this check because a language model may generate numbers that look correct but don’t reconcile.
Real receipts also obey a template. A given merchant prints from a handful of point-of-sale configurations, with a recognizable store number format, terminal identifier, timestamp convention and footer.
When you hold millions of receipts from that merchant, a fabricated one can stand out beside genuine ones. It is rarely wrong in any single place. It is wrong relative to the entire population.
Then there is corroboration.
Card data, travel itineraries, calendar entries and colleagues’ expenses on the same trip either support the transaction or they don’t. A dinner receipt that never ran through a card feed can be a stronger signal than a font that looks slightly off.
Employee expense behavior closes the loop. One recurring pattern we see is repetition: one underlying image, with the amount and date edited, submitted just outside a duplicate-check window.
Layered Reviews are the Best Answer to AI-Generated Fraud
No single test of a receipt is reliable enough to act on by itself.
A forger can strip the metadata from an image and ensure the arithmetic adds up. Intent varies, too, because some fabricated receipts may simply be employees reconstructing genuine expenses after losing the original.
Detection must combine the full set of weak signals, score them together and provide an explanation.
A system that returns a confidence score and the full audit context lets a human reviewer investigate and determine intent. A system that returns only a pass or fail can force auditors to guess, introducing human bias and error.
Three Changes Organizations Should Make Today
Making these three changes before the next quarter closes can improve the chances of catching an AI-generated receipt.
- Audit every report before payment. Most enterprises review 10% to 20% of expense reports, selected by rules or sampling. Partial coverage creates exposure.
- Write down the policy. Many expense policies never state whether recreating or manipulating receipts is permitted. Make it clear so intent is handled consistently.
- Corroborate instead of inspect. Any control that depends on a person recognizing a fake by eye is already becoming obsolete.
Receipt images will keep getting easier to fake, but there will always be evidence to gather.
A real dinner leaves a card authorization, a calendar entry, a merchant record that matches thousands of others and sometimes a colleague who expensed the same meal. Across the receipts we see at AppZen, an AI-generated image might pass visual inspection but fall apart everywhere else.
That is where finance teams should be looking.
An AI model can generate a picture of a dinner. It still cannot generate the dinner itself.
Kunal Verma is CTO and Co-founder of AppZen, where he has spent more than a decade building AI for financial document forensics.
