A fake receipt is a fabricated or altered document submitted as proof of a business expense that did not happen, or did not happen as the claim describes. Visual review no longer detects one reliably. An auditor catches one with five checks, because each tests the document against facts outside the image.
Producing a convincing forgery used to take either a template site with recognizable output or real effort in an image editor. Both left artifacts a trained reviewer learned to spot. Image models removed that tell within about eighteen months. Detection therefore depends on the facts around the image rather than on the image itself. Those are whether the merchant exists, whether the arithmetic adds up, whether the price is plausible, and whether the same document has already been paid.
A current image model produces a receipt with correct merchant branding, plausible item pricing, realistic paper wear, a believable fold, and handwritten tip lines in about half a minute. Nothing about the resulting image reads as synthetic to a human reviewer working through a queue.
Submission data shows how quickly this took hold. PYMNTS reported our platform data in 2026, in coverage of AI-generated fake receipts. Generated documents went from zero percent of flagged fraudulent receipts in March 2025 to 70.8 percent by mid-May 2026. That sample covered 1,471 fake receipts submitted by 745 employees at 174 companies, worth $148,143 in fabricated reimbursements. The average fake came in around $100, below the $182 average of the older template-based fakes.
That drop in average value is the detail worth pausing on. Fabricating a receipt got easy enough that people stopped saving it for expenses large enough to justify the risk. HR Executive reported a 2026 employee AI receipt survey of 2,000 workers in the United States and the United Kingdom. Four in ten US employees had used AI to create a fake receipt on a business expense report. Nearly 20 percent fabricated a purchase that never happened, about 15 percent inflated a real one, and 6 percent replaced a receipt they had lost.
Manual receipt review asks a person to answer a question they have no evidence for. Looking at an image shows whether the document appears well-formed. It says nothing about whether the restaurant exists or serves that dish. It says nothing about whether the tax rate matches the jurisdiction, or whether the same meal already cleared on a colleague's report last month.
Three structural problems compound that.
Reviewers work under time pressure, and they lose attention across a queue. The fiftieth receipt of the morning receives less scrutiny than the fifth.
Coverage is partial. Most enterprises audit 10 to 20 percent of expense transactions, a gap examined in our analysis of 100 percent audit coverage. The majority of submissions therefore receive no review at all, and a forgery only needs to land in the unreviewed portion.
Review often happens after reimbursement. The Association of Certified Fraud Examiners studied 2,402 cases across 143 countries and territories for Occupational Fraud 2026 and found a median scheme length of 12 months before detection. The median loss was $104,000 per case. Schemes caught inside six months showed a median loss of $40,000. The loss depends more on how soon a team catches a scheme than on how accurately it detects one.
Detection works when the document is tested against facts outside the image. These five checks, run together on every submission, catch the great majority of forgeries.
Generated and edited images contain metadata trails. Generator sites leave identifiable signatures, editing tools record their own, and compression history differs from a photographed paper receipt. Provenance analysis flags documents whose origin does not match a phone camera.
Confirm the merchant exists, operates at the claimed address, was open on the claimed date, and sells the items listed. A hotel that does not serve a buffet cannot produce a buffet charge. This check rules out fabricated merchants outright, and it also catches real merchants used as cover for invented purchases.
Line items, modifiers, discounts, tax, and tip have to sum to the stated total at the tax rate for that jurisdiction. Image models produce plausible numbers rather than correct ones, so arithmetic failure is one of the most reliable signals available.
Compare the price against known pricing for that merchant and category. A $65 omelet at a diner fails on economics regardless of how clean the document looks.
Check the document against every receipt the organization has already processed, including submissions from other employees, other periods, and other systems. Duplicates, resubmitted originals with edited dates, and shared forgeries surface here and nowhere else.
Most expense platforms added receipt validation as a rules layer, checking that an attachment exists, that it is legible, and that the total matches the entered amount. Those checks confirm a receipt was uploaded. They do not test whether it is true.
A second group added detection of AI-generated imagery as a standalone feature. A standalone classifier degrades quickly, because image models improve faster than the classifiers trained on their previous output. Any control resting on a single signal stops working as soon as image models improve again. Merchant reality and arithmetic do not expire, because a forged receipt has to describe a transaction that did not happen.
Our AI runs all five checks on every receipt on every report before reimbursement. The models behind them were trained across millions of real and fraudulent receipts collected over a decade of enterprise expense data. Provenance, merchant verification, arithmetic, pricing, and cross-report history are evaluated in one pass, so a document that survives one check still has to survive the others.
Coverage is what makes the approach work. Because the audit runs on 100 percent of submissions rather than a sample, there is no unreviewed portion for a forgery to land in. For the practitioner view of how this plays out day to day, read our AI-fake receipts defensive playbook, or the research behind detecting AI-generated fake receipts.
A flagged document is not proof of intent. It is evidence that a receipt failed a named check, and a genuine receipt recreated after the original was lost looks the same. Query first, escalate on repetition, and keep the judgment with the review team.
Pull a sample of approved receipts from the last quarter and test them against merchant reality and arithmetic rather than appearance. Then compare what you find against what the current process flagged. The gap between those two numbers is the real exposure. Our AI expense audit page describes how that gap closes before payment.
A fake receipt is a fabricated or altered document submitted as evidence of a business expense that never occurred, occurred at a different amount, or was already claimed. It includes fully generated images, edited originals, and duplicated documents with changed details.
Enterprise expense data reported in 2026 found AI-generated documents making up 70.8 percent of flagged fraudulent receipts by mid-May 2026, up from zero percent in March 2025. A separate 2026 survey of US and UK workers found four in ten US employees had used AI to create one.
Visual inspection no longer works reliably. Current image models reproduce merchant branding, paper wear, and handwriting convincingly. Detection has to test the document against merchant records, tax arithmetic, pricing, and submission history.
Consequences vary by employer and jurisdiction, and they typically range from repayment and disciplinary action to termination. Fabricating expense documentation may also constitute fraud under local law, so legal counsel should be involved before a finding is characterized.
Audit every report before reimbursement rather than sampling after it, and test receipts against external facts instead of appearance. Route exceptions by risk so reviewers spend their time on the claims with the most money at stake.