> ## Content Index > Fetch the complete content index at: https://fintekcafe.com/llms.txt > Use this file to discover other available public pages before exploring further. # AI-Generated Fake Receipts: How Expense Fraud Changed and How Finance Teams Catch It - URL: https://fintekcafe.com/ai-fake-receipts-expense-fraud-explained/ - Published: 2026-10-23T09:00:00.000Z - Updated: 2026-10-23T09:00:00.000Z - Description: AI-generated fake receipts now beat visual review. How expense fraud changed, which detection still works, and the controls to use by company size. - Author: FinTekCafe - Tags: finance, ai, fraud, expenses, seo **An AI-generated fake receipt is a receipt image produced by an image model from a text prompt, for a purchase that never happened or at a price that was never paid. It has no smudged digits, mismatched fonts or pixelated logos, so the visual review that expense controls were built on no longer catches it.** The shift has been fast. Data from AppZen, which audits expense reports for large enterprises, published in June 2026, showed AI-generated images made up 70.8 percent of the fraudulent receipts it flagged by mid-May 2026, up from zero in March 2025, the month OpenAI put a far more capable image generator into ChatGPT. The position taken here: stop treating the receipt as evidence and treat it as a claim to check against data the employee cannot generate, such as the card transaction, the booking and the calendar. Detection tools help at the margin. The structural fix is to move spend onto rails where the receipt no longer carries the trust. ![Which schemes](https://fintekcafe.com/content/images/2026/10/inline-01-which-schemes.png) ## Key Takeaways - Expense fraud is old, common and slow to detect: the ACFE's 2026 report found it in 13 percent of occupational fraud cases, running a median 18 months. - AI changed the cost of forgery, not the motive. A convincing receipt now takes a sentence, so visual review is no longer a reliable control. - The size of the shift is reported mainly by vendors who sell detection, and figures vary by population. Treat them as direction, not precise measurement. - Metadata and watermark checks catch careless fraud, but screenshots and format conversion can erase or weaken their signal. Matching claims against card feeds, bookings and cross-employee patterns still works, because the forger does not control that data. - The structural answer is to remove the receipt from the trust chain: cards for most spend, per diems for meals, and pre-approval for anything large, leaving a small, scrutinized pool of out-of-pocket claims. ![Detection methods](https://fintekcafe.com/content/images/2026/10/inline-02-detection-methods.png) ## How Expense Fraud Works, and What It Costs **Expense reimbursement fraud is an employee claiming money for business expenses that were fictitious, inflated or personal.** The ACFE's Occupational Fraud 2026: A Report to the Nations, published in May 2026 and based on 2,402 cases investigated between January 2024 and September 2025 across 143 countries and territories, gives the best independent picture of the scheme's economics: - **Frequency.** Expense reimbursement schemes appeared in 13 percent of all cases (306 cases). - **Size.** The median loss was 36,000 dollars per case. - **Duration.** The median scheme ran for 18 months before it was detected, a slow drip of around 1,900 dollars a month. - **Who is most exposed.** Small organizations, those with fewer than 100 employees, saw expense schemes in 19 percent of their cases, against 10 percent at larger organizations. Two other findings matter for what follows. Tips were the most common way any occupational fraud came to light, in 43 percent of cases. And organizations that used proactive data monitoring and analysis had a median loss 53 percent lower than organizations without it (70,000 against 150,000 dollars) and detected fraud 44 percent faster (9 months against 16). That second finding is the quiet thesis of this article: the controls that work are the ones that analyze data, not the ones that look at documents. ![Shrink pool](https://fintekcafe.com/content/images/2026/10/inline-03-shrink-pool.png) ## What Changed: Forgery Became Free ### The capability shift Fake receipts are not new: online receipt generators and editable templates have existed for years. What changed in 2025 is that general-purpose image models became good enough to produce photorealistic documents on request. On March 25, 2025, OpenAI replaced the image generator in ChatGPT with native GPT-4o image generation, announced for all tiers, reaching free users about a week later. Within days, expense platforms were responding: Ramp released a tool to flag AI-generated receipts on April 3, 2025, explicitly in response to that update. The skill and cost needed to forge a receipt fell to roughly zero. A prompt can specify the restaurant, date, itemized dishes, tax line and tip, and ask for a creased, phone-photo look. SAP Concur, writing in December 2025, described AI-generated receipts as "so realistic they are evading traditional human audits", with "plausible logos, fonts, itemized entries, VAT lines, and totals." ### What the reported numbers say The evidence on scale comes largely from vendors with a product to sell, so read each figure with its population in mind. | Source and date | What it reported | Read it as | | ------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------- | | AppZen, via the Financial Times (October 2025) | About 14 percent of fraudulent documents flagged in September 2025 were AI-generated, against none a year earlier | Early signal on large-enterprise expense audits | | AppZen, reported by Accounting Today and PYMNTS (June 2026) | 70.8 percent of flagged fake receipts AI-generated by mid-May 2026; 1,471 AI receipts from 745 employees at 174 companies, totaling about 148,000 dollars | AI has displaced templates as the forger's tool of choice | | Medius survey by Censuswide (July 2025, 1,000 US and UK finance professionals) | 30 percent saw a rise in faked receipts; 32 percent said they would not recognize a fake | Finance teams feel the shift and doubt their eyes | | Emburse survey by Atomik Research (June 2026, 2,000 US and UK workers) | 40 percent of US and 29 percent of UK respondents said they had used AI to generate a fake receipt | Self-reported, and includes recreating lost receipts for real purchases | | SAP Concur (December 2025) | About 1 percent of reviewed receipts flagged as potentially AI-generated | Most receipts are still genuine | | Inscribe, 2026 State of Document Fraud Report (January 2026) | AI-generated documents under 5 percent of fraudulent documents in its lending network in 2025, though up nearly fivefold from May to December | In other document populations, templates still dominate | Two details in the AppZen data matter. The AI-generated fakes were small: roughly 100 dollars on average, against 182 dollars for the template fakes they replaced, with a median of about 32 dollars. Small claims sit under auto-approval thresholds and rarely attract an auditor's sample. The cost so far is not large single losses. It is the collapse of a control organizations assumed was working. SAP Concur offered the most useful framing: AI "isn't increasing the frequency of expense fraud, but it is changing how it occurs." The method has moved to a place where old controls cannot see it. ## The Fraud Schemes, and How AI Changes Each One AI changes some expense schemes far more than others. The table maps each to the control that still catches it. | Scheme | How AI changes it | Control that still catches it | | ----------------------------------------------------------------- | ----------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------- | | Fictitious expense (a purchase that never happened) | Large change: a realistic receipt for any merchant, date and amount on demand | No matching card transaction; merchant or location inconsistent with travel records; employee pattern analytics | | Inflated expense (real purchase, higher amount) | Large change: regenerating a genuine receipt with a new total is trivial | Card-feed amount match; merchant-category price norms; comparison with colleagues on the same trip | | Disguised personal expense (real receipt, false business purpose) | Small change: the receipt is genuine, only the purpose is false | Calendar and booking cross-checks; purpose review for high-risk categories; pattern analytics | | Altered line items (hiding alcohol or banned items) | Large change: a clean itemized receipt without the disallowed lines | Itemized data from the card network or merchant where available; category anomaly scoring | | Duplicate submission | Moderate change: the same expense regenerated as a different-looking image | Duplicate detection on transaction data rather than image similarity | | Recreated lost receipt (genuine purchase) | New gray area: an employee recreates a receipt they lost | Clear policy offering a missing-receipt declaration instead, so honest employees have a legitimate path | | Mileage and per diem padding | Little change: no receipt involved | Route and distance validation; per diem date checks against travel bookings | | Corporate card misuse (personal spend on a company card) | Little change: the receipt is not the evidence | Merchant category blocks, spend limits, prompt transaction review | AI damages controls that rely on the receipt as proof. It does little to controls that rely on data the employee does not create. That is the basis for every recommendation that follows. ## Detection That Still Works, and Its Limits ### Card-feed and merchant data matching **The single strongest detection control is matching each claim against an independent record of the transaction.** On a corporate card, the card feed records the merchant, amount, date and often location, so the receipt becomes supporting evidence for an expense the organization already knows happened, and forging it is mostly pointless. The logic is the same as in [matching bank transactions to internal records](https://fintekcafe.com/payment-reconciliation-explained/): reconcile the claim against a source the claimant does not control. The limit is coverage. Card matching does nothing for out-of-pocket claims, which is where AI fakes concentrate. ### Metadata and provenance checks Images produced by some generators carry provenance signals. OpenAI's documentation states that supported images from ChatGPT, Codex and the OpenAI API carry C2PA Content Credentials, signed metadata identifying AI use, and, since OpenAI announced it on May 19, 2026, a SynthID watermark embedded in the image itself; most receipts in AppZen's twelve-month data predate it. Ramp's April 2025 tool and SAP Concur's AI-generated receipt checker both inspect metadata for generation markers, and SAP Concur reported that its checker flags about 18 times more suspect receipts than its earlier checks for known online receipt generators. The limits are stated by OpenAI itself: compression, cropping, screenshots, metadata removal and format conversion can erase or weaken these signals, and not all OpenAI-generated content carries one. Ramp described its tool as a first line of defense against casual fraud, not a complete answer. Metadata checks are cheap and catch the careless, who are a real share of offenders. They are not a control to rely on. ### Duplicate and pattern analytics Pattern analytics look across claims rather than at one image: repeated claims just under the receipt threshold, claims on days the travel system shows the employee at home, a merchant that appears for one employee and nobody else, or meals far above what colleagues on the same trip claimed. These methods work because they do not depend on the image at all. They are the expense-side equivalent of the transaction scoring described in [how payments fraud detection moved to milliseconds](https://fintekcafe.com/real-time-fraud-detection-payments-2026/): signals from behavior, scored against a baseline. This is also the control the ACFE evidence supports most strongly. Proactive data monitoring was associated with about half the median loss and 44 percent faster detection, yet only 17 percent of organizations with fewer than 100 employees had it, against 58 percent of larger ones. ### Vendor AI detectors Most expense platforms now sell an AI layer that scores receipt images for signs of generation. These detectors are useful and improving, but their claims are vendor claims, for three reasons. Detection rates are self-reported on the vendor's own population. Generators improve faster than detectors are retrained. And false positives mean a genuine employee with a crumpled real receipt gets questioned, a cost to trust that rarely appears in the business case. The right buying question is not "what percentage of AI receipts do you catch?" but "what do you check that the employee cannot fake?" Card data, booking data and cross-employee analytics are a durable control. Image forensics alone is an arms race. ## The Structural Fix: Take the Receipt Out of the Trust Chain ### Corporate and virtual cards The most effective response to receipt forgery is to need fewer receipts as evidence. A corporate card gives the organization the transaction before any receipt arrives. A single-use or merchant-locked virtual card constrains the spend before it happens: a set amount, merchant and period. The mechanics are covered in [how single-use virtual card numbers fix corporate spend](https://fintekcafe.com/virtual-cards-b2b-payments/), and they apply to employee spend as much as to supplier payments. Cards do not remove fraud. They move it to card misuse, which is easier to see because the data arrives in real time and can be limited by merchant category. ### Per diems for meals and incidentals Meals are a high-volume, low-value category where receipt audits are least worth the effort. A per diem, a fixed daily allowance for meals and incidentals, removes the receipt entirely. In the United States, IRS Publication 463 (2025 edition) states that documentary evidence is not needed for meals or lodging accounted for under an accountable plan using a per diem allowance method, and that documentary evidence is also not needed for expenses other than lodging of less than 75 dollars. Those are tax substantiation rules, not fraud controls, and an organization may set stricter internal policy. Outside the United States, per diem treatment differs by jurisdiction and should be confirmed with tax advisers. A per diem will sometimes pay more than an employee spent. For most companies, that occasional overpayment costs less than reviewing thousands of meal receipts that can now be forged in seconds. ### Pre-approval for large or unusual spend Above a threshold, the claim should not be the first time finance learns about the expense. Pre-approval creates a record before the spend for the claim to match. Where a supplier is involved, the purchase belongs in the procure-to-pay process with its own controls, as set out in [what AI can take over in invoice-to-pay](https://fintekcafe.com/accounts-payable-automation-ai-explained/), not in an expense report. ### What is left After cards, per diems and pre-approval, what remains is a smaller pool of genuine out-of-pocket claims: a taxi in a cash-only city, an emergency purchase, mileage. That is where receipt scrutiny and AI detectors should concentrate. A control applied to a small share of claims can be applied properly; one spread across all claims is applied by skim-reading. ## The People Side ### Policy and consequence Expense fraud is shaped by norms. In the Medius survey, 29 percent of finance professionals admitted bending expense rules themselves, and 34 percent said they had been pressured to approve an expense that did not seem legitimate. Three policy elements matter most: - **State plainly that generating or altering a receipt is fraud,** including "recreating" a lost one. Offer a missing-receipt declaration as the legitimate alternative, so honest employees are never tempted into a gray area. - **Apply a known, consistent consequence,** regardless of seniority. - **Tell employees what is checked.** Knowing claims are matched to card and travel data is itself a deterrent. ### Reimbursement speed The Emburse survey reported that 51 percent of US respondents who admitted submitting fraudulent claims had incurred overdraft fees, late payments or credit card interest while waiting to be reimbursed. Slow reimbursement does not justify fraud, but it breeds resentment. Paying valid claims quickly, and moving routine spend onto company cards, reduces both pressure and rationalization. ### Not treating every employee as a suspect SAP Concur flagged about 1 percent of reviewed receipts as potentially AI-generated, and most expense claims are honest. Forensic review of every lunch costs more in time and goodwill than it saves. The aim is a system the honest majority never notices, because their spend arrives through data the organization already trusts, with scrutiny reserved for claims and patterns that do not fit. ## A Decision Framework by Company Size and Spend Profile The right control stack depends on how many people spend company money and how much of that spend happens outside a company card. | Company size | Typical spend profile | Minimum control stack | Add when out-of-pocket share is high | | ------------------------- | ------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------- | | Under 100 employees | Few spenders, owner or one finance person reviews everything | Corporate cards for anyone who travels; per diem for meals; written policy; owner reviews card feed monthly | Expense tool with built-in metadata and duplicate checks; second reviewer for the reviewer's own claims | | 100 to 1,000 employees | Sales and field teams travel; managers approve | Cards for all regular travelers; card-feed matching; per diems; pre-approval above a threshold; hotline or simple reporting channel | Pattern analytics in the expense platform; quarterly targeted audit of high out-of-pocket claimants | | 1,000 to 10,000 employees | Many spenders, several countries, mixed card and cash | Full card program with merchant category controls; virtual cards for events and bookings; automated policy and duplicate checks; vendor AI detection on out-of-pocket receipts | Cross-employee analytics; travel and calendar data integration; dedicated expense audit function | | Over 10,000 employees | Global, high volume, varied local rules | All of the above, plus continuous audit of 100 percent of claims by software and human review only of flagged items | Country-specific per diem and receipt rules; fraud data shared with internal audit and HR | Spend profile matters as much as size. A firm whose people mostly fly, stay in booked hotels and eat with clients can move almost all spend onto cards and per diems. A firm with field technicians buying parts, drivers paying tolls in cash, or staff in markets where cards are not widely accepted has a large out-of-pocket pool, and should invest more in pattern analytics and targeted review. Field service, construction and many healthcare organizations sit in this second group, whatever their size. ### The position, and its tradeoff **The recommended stance is to stop investing in better ways to look at receipts and start investing in ways to need them less.** Move every recurring category of spend onto cards, per diems or pre-approved bookings, and concentrate detection on the small remainder. The tradeoff is cost and control. A card program carries administration and the risk of card misuse. Per diems sometimes overpay. Some managers will resist losing the flexibility of paying personally and claiming later. And detectors are still needed for what remains, so the decision is about which approach carries the load. The alternative, an escalating contest between image generators and image detectors, is one finance teams will not win, because generators improve faster than detectors are retrained and the forger only needs to win once per claim. ## Frequently Asked Questions ### Can AI-generated fake receipts be detected? Some can. Images from major generators may carry provenance metadata or watermarks, but screenshots, cropping and format conversion can erase or weaken those signals. The more reliable detection matches the claim against data the employee cannot fake: the card transaction, the travel booking and spending patterns across colleagues. ### How fast are AI-generated receipts growing? AppZen data published in June 2026 showed AI-generated images made up 70.8 percent of the fake receipts it flagged by mid-May 2026, up from zero in March 2025\. Other populations show a smaller share: Inscribe found AI-generated documents were under 5 percent of fraudulent documents in lending during 2025\. The direction is clear even though the precise share depends on who is measuring. ### What is the best way to prevent expense fraud? Reduce the number of claims that depend on a receipt as proof. Cards supply the transaction data directly, per diems remove meal receipts, and pre-approval creates a record before large spend. Detection then concentrates on the smaller pool of out-of-pocket claims. ### Is an employee recreating a lost receipt with AI committing fraud? Under most policies, yes, because the document submitted is not a genuine record of the transaction, even if the purchase was real. The practical fix is a clear missing-receipt declaration process, so employees with a genuine lost receipt have a legitimate way to claim without creating a document. ## The Bottom Line AI did not invent expense fraud, and it has not yet made it much more expensive per claim. What it has done is end the usefulness of the receipt as proof. The expense report rested on an assumption that a receipt was hard enough to fake that looking at it was a reasonable check. That assumption failed in 2025, and every control that depends on someone looking at an image is now weaker than it appears on paper. The organizations that handle this well will not have the cleverest image detector. They will have moved most spend onto cards, per diems and pre-approved bookings, matched every remaining claim against data the employee did not create, and told their people plainly what is checked. That approach catches AI-generated receipts as a side effect, and will still work when the next generation of image models arrives.