Field Staff Photograph Their Receipts, and AI Logs Them in a Sheet

Field staff pay for fuel, parking, client meals, or small materials, then send a photo of the receipt to a WhatsApp group. At month-end, finance scrolls through hundreds of photos, retypes each date and amount, then matches them against the money already paid out. Some receipts are blurry, some were sent twice, some were never sent.

AI can now help with this: the receipt photo gets read, its contents go into a sheet, and finance only has to check and approve. But there's a new risk to understand first: fake receipts made by AI.

The flow from photo to payment

1. 1Receipt photoSent through a form or bot, not a group chat
2. 2AI reads itDate, store, total, and category
3. 3Into the sheetOne row per receipt, with a link to the photo
4. 4Automatic checksDuplicates, odd amounts, missing data
5. 5Manager approvesOnly then is it paid

The key is the first step: receipts shouldn't go to a group chat mixed with other conversation. Use one dedicated channel, such as a form or a WhatsApp number connected to the system, so every photo is logged immediately with who sent it, when, and for which project.

The AI then reads the receipt: the date, store name, total, and if possible the category. Faded thermal-printer receipts and handwritten notes are often read less accurately, so the system should mark low-confidence rows or rows with missing data as "needs checking" instead of guessing.

A new risk, fake receipts made by AI

In March 2025, TechCrunch reported that ChatGPT's new image generator was already being used to make convincing fake restaurant receipts. A year later, the data started showing up inside companies.

Forbes (June 2026) cited data from AppZen, a company that audits expense claims for large enterprises. AI-generated receipts went from 0% of fake receipt flags in March 2025 to 70.8% by mid-May 2026. In the 12 months to May 15, 2026, AppZen found 1,471 AI-generated receipts from 745 employees at 174 companies, with total claims of US$148,143. Those are detected claims, not confirmed losses.

Share of AI-made receipts among AppZen's fake receipt flags, mid-May 2026

70.8%

Median AI-made receipt, kept small to slip through without review

US$32

AI-made receipts detected in 12 months, from 745 employees

1,471

The most interesting part is the amounts. In the same data, AI-made receipts averaged US$101 with a median of about US$32, smaller than old-style fake template receipts, which averaged about US$182. The pattern isn't one big claim but many small ones that slip under auto-approval limits. This data comes from large companies abroad, but the pattern applies to any business with a rule like "anything under a certain amount gets paid automatically".

Sensible safeguards

Forbes suggests companies verify transactions against several data sources, adjust auto-approval limits, and combine several kinds of checks. Translated for smaller businesses, here's what I'd suggest:

1. Check for duplicates automatically. The same date, store, and total from the same person gets flagged right away.
2. Match against other sources. Company card or e-wallet statements, visit schedules, or delivery routes. A fake receipt has a hard time getting through when the transaction has to exist in another record.
3. Don't set an "auto-pay" limit that's too loose. If you keep a limit, take a random sample of claims below it for review.
4. For routine spending, use a company card or e-wallet so there's a transaction record, not just a receipt photo.
5. A manager still approves before payment. Money going out is an action that needs human approval, per the levels in At What Point Should Humans Still Check AI and Automation Results.

Start simple

You don't need to buy expense management software right away. For small to mid-sized teams, the flow above can be built with a form or WhatsApp bot, Google Drive for the photos, Google Sheets for the log, and an AI model to read the receipts. When Sheets is still enough and when to move on is covered in When Google Sheets Is Enough as a Web App Database, and When to Move On.

The same "proof comes in, a human checks it, then it's processed" pattern is what I used in the co-living finance system, where tenants' payment proofs are checked by the management team before the receipt goes out. And if the monthly summary later needs to go to many people at once, how to do that is in Making Thousands of Certificates, Invitations or Invoices From One Sheet.

Wrapping up

AI removes the most tedious part, retyping receipts. But AI also makes fake receipts cheap. So the automation has to come with better checks: duplicate detection, matching against other sources, and a human approving money going out.

If your field team still sends receipts through a group chat and finance still logs them by hand, tell me about the flow first.