A fake receipt is a document submitted as proof of a business expense that the claim misrepresents. The purchase never happened, the amount differs from what was paid, or the document was recreated after the original went missing. Visual review no longer detects them. An auditor catches them with six checks, and all six read evidence outside the picture.
Key takeaways
- AI-generated documents rose from zero to 70.8 percent of flagged fraudulent receipts between March 2025 and May 2026, across 1,471 receipts from 745 employees at 174 companies.
- Value per fake fell as volume rose, from $182 for template fakes to about $100 for AI-generated ones. Forgery became cheap enough to use on small claims.
- Reconciliation against the corporate card feed is the highest-yield check and the one most programs skip. No matching transaction means the image quality is irrelevant.
- Not every fake receipt is fraud. About 6 percent of surveyed workers used AI to replace a lost receipt for an expense that genuinely happened.
Until recently these documents came from a short list of template websites, and a trained reviewer could identify one on sight. That tell is gone.
How fake receipts changed, and how quickly
The shift took about fourteen months. PYMNTS reported enterprise expense platform data in June 2026. It showed AI-generated documents rising from zero percent of flagged fraudulent receipts in March 2025 to 70.8 percent by mid-May 2026. The sample covered 1,471 fake receipts submitted by 745 employees across 174 companies, worth $148,143 in fabricated reimbursements.
Value per fake fell while volume rose. The average AI-generated fake came in around $100, against $182 for the older template-based fakes. Forgery became cheap enough that submitters stopped reserving it for claims large enough to justify the effort.
Employee self-reporting matches the platform data. A 2026 survey of 2,000 US and UK workers reported by HR Executive found that four in ten US employees have used AI to create a fake receipt. Nearly 20 percent fabricated a purchase that never happened, about 15 percent inflated a real one, and roughly 6 percent used AI to replace a lost receipt for an actual expense. That last group matters for control design, because a policy treating every replaced receipt as fraud produces more aggrieved employees than recoveries.
The cost of missing the pattern compounds with time. The Association of Certified Fraud Examiners (ACFE) studied 2,402 cases for Occupational Fraud 2026. Median loss ran to $104,000 per case, with a median scheme duration of 12 months before detection. Asset misappropriation, the category expense fraud belongs to, appeared in 90 percent of cases. Schemes caught within six months showed a median loss of $40,000, while schemes lasting five years or longer exceeded $1.1 million. Expense fraud repeats on a monthly cycle, so every month a pattern goes undetected is another entry on that curve.
Why the image is the wrong place to look
Wolfgang Beltracchi sold forged modernist paintings for decades without an expert eye catching him. A chemist did, by identifying a pigment that had not been manufactured when the painting was supposedly made. The forgery was flawless as an image and false as an object.
Fake receipts now behave the same way. Because the image is convincing, the image is the wrong place to look.
The working principle is to treat a receipt as a claim rather than as proof. A claim can be tested against evidence the organization already holds and against the physical world it describes. A restaurant that closed in 2024 cannot have served dinner last Tuesday. A $19 breakfast at a hotel with a published $14 breakfast menu is a discrepancy worth questioning. A card feed showing no transaction at that merchant on that date settles the matter without any image analysis at all.
Timing is the second principle. A control that runs after reimbursement produces a collections problem, and collecting from current employees is a conversation most finance teams avoid. Every check below is worth more before the payment file is cut than after it.
Proportion is the third principle. Most of what these checks surface is not fraud. It is a lost receipt someone replaced, a personal item on a business tab, a duplicate submitted across two cycles, or a rounding error nobody noticed. Those findings are recoverable, and they call for a lighter response than a fraud investigation.
The sequence of checks for fake receipts, cheapest first
Six checks follow, arranged so the inexpensive ones run first and the expensive ones run only on the claims the earlier checks did not clear.
1. Read the file behind the picture
Examine the image file itself, not only what it depicts. Every image carries structure the eye does not see. Generation software leaves signatures in the file, compression history shows whether an image was saved once or edited and resaved, and color distribution in a synthetic image differs from a phone photograph of paper under real light. Two receipts produced by the same model from the same prompt share characteristics that no two genuine photographs share.
This check misses a fake that was printed and then photographed, because printing resets much of the file evidence. It also misreads a genuine receipt captured as a screenshot, which looks synthetic and is not.
2. Confirm the merchant exists and charges those prices
Verify that the merchant is real, that it operated at that address on that date, and that its category matches the claim. Then test the prices against what that merchant actually charges. The strongest signals are outside the document. Beltracchi's pigment was an anachronism, a $65 omelet is the same class of error, and so is a taxi fare implying a speed no city allows.
3. Test the arithmetic
Add the line items and compare the result to the stated subtotal, then check the tax against the local rate and the tip against the total. Generated receipts are visually correct and mathematically careless. Line items that do not sum to the subtotal are common in fabricated documents, as is tax that misses the jurisdiction rate. The check costs almost nothing to run and catches a meaningful share of fakes on its own.
4. Check what a genuine receipt would contain
Compare the document against the fields a real receipt from that merchant carries. Terminal identifiers, sequential transaction numbers, tax registration numbers, and the last four digits of the card all appear in predictable places. Fabricated receipts tend to include what looks important and omit what looks like clutter, so a missing field is a signal.
5. Reconcile against data you already hold
Match the receipt to the corporate card feed, then against prior submissions from the same employee, against submissions from colleagues on the same trip, and against the booked itinerary. This check has the highest yield and most programs skip it. Duplicate submissions across reporting cycles are the clearest example, and duplicate spend found rises sharply once a full year of history is in place.
6. Score the submitter over time
Retain the signals from every check, including the ones that cleared, and track them by person. One suspect receipt is an incident. Four suspect receipts from one person across six months is a pattern, and the pattern is what supports an investigation.
Where fake receipt programs fall short
Most programs fail on coverage rather than on technique. Teams running a sample-based audit review 10 to 20 percent of expense reports, which leaves a forged receipt a four in five chance of never being examined. Auditing 100 percent of expense reports turns these checks from a spot inspection into a control.
The second failure is single-signal thinking. A team buys an image classifier and treats its score as a verdict, then either drowns in false positives or loses trust in the tool the first time it clears a fake. A classifier is one input among six.
The third failure is timing. Manual receipt verification runs after reimbursement in most organizations, and recoveries after payment are a fraction of prevented payments.
The fourth failure is human capacity. Reviewers working a queue apply the checks unevenly, spend their attention on low-risk reports, and get slower as volume grows. That is arithmetic rather than training.
How we approach fake receipts
Our AI reads every line of every receipt on every report before reimbursement, across 100 percent of expense reports rather than a sample. Audits run in 42 languages across 97 countries. That range matters because international receipts, including the fapiao issued as an official tax invoice in China and value-added tax (VAT) documents, are subject to validation rules a generalist reviewer will not apply consistently.
The six checks above run together rather than in isolation. Image provenance and metadata, pattern recognition, merchant authentication, mathematical validation, and completeness verification each contribute to a single disposition, so the outcome never depends on one signal. Our detail on the five detection layers sets out how each one is built.
Clean reports auto-approve, and the platform routes exceptions to a human with the evidence attached, so the reviewer starts from a finding rather than from a blank queue. Customers reach auto-approval rates above 75 percent and remove 80 to 90 percent of manual audits, which puts auditor attention on the reports that need judgment.
The bottom line
Fake receipts are now a volume problem rather than a craft problem, and a team answers a volume problem with coverage plus corroboration. Identify the two checks your process does not run today, most often the card-feed reconciliation and the merchant price test, and add them before payment rather than buying a tool first. Our AI expense audit overview describes what full pre-payment coverage involves.
Frequently asked questions
What are fake receipts?
Fake receipts are fabricated or altered documents submitted as proof of a business expense that did not occur, or did not occur as the claim describes. They include fully invented receipts, genuine receipts with amended amounts, and replaced versions of receipts an employee lost.
How common are AI-generated fake receipts?
Platform data reported by PYMNTS in June 2026 found that AI-generated documents made up 70.8 percent of flagged fraudulent receipts by mid-May 2026, up from zero percent in March 2025. A 2026 survey of 2,000 workers found that four in ten US employees admitted using AI to create a fake receipt.
Can a person still spot a fake receipt by looking at it?
No. Current image models reproduce merchant branding, paper texture, folds, and handwriting well enough that visual review is no longer a reliable control. Detection now depends on file-level evidence, merchant verification, arithmetic, and reconciliation against data the company already holds.
What is the single highest-yield check?
Reconciliation against the corporate card feed. Where no matching transaction exists at that merchant on that date, the quality of the image is irrelevant, and the finding needs no forensic argument to support it.
Does every fake receipt indicate fraud?
No. About 6 percent of employees in the 2026 survey said they used AI to replace a lost receipt for a purchase that genuinely happened. Policy should separate a fabricated purchase from a replaced document, because the appropriate response differs.