How expense platforms can detect fake receipts before they're processed

Expense fraud is one of the most persistent and underestimated problems in corporate finance. The ACFE estimates that organisations lose approximately 5% of annual revenue to occupational fraud, with expense reimbursement fraud among the most pervasive categories. Yet despite this, most expense platforms validate receipts almost entirely through OCR-based text extraction and rule checking — a stack that was never designed to detect pixel-level manipulation.

The gap has widened in the last two years as AI editing tools have made receipt manipulation accessible to anyone. Changing a restaurant bill from $34 to $340 now takes under two minutes and produces a result that passes every standard check. The only evidence of the manipulation is in the pixel data — which most expense platforms never examine.

What expense fraud looks like in practice

Amount inflation is the most common form of expense fraud and the hardest to catch with traditional methods. The fraudster takes a legitimate receipt, edits the amount upward, and submits the modified image. The document passes all validation: the vendor is real, the date is valid, the amount is within policy limits, and OCR extracts the inflated value without any indication it has been changed.

Why rule-based checks fail

Most expense platforms check for: amount within policy limits, vendor legitimacy, date within submission window, and duplicate invoice numbers. A fraudster who knows your rules can comply with all of them while still submitting a fake document. Amount inflation is the clearest example — the inflated amount is within policy limits by design.

Fabricated receipts are harder still. A fraudster who generates a fake receipt from a real restaurant — correct name, address, and business format — will pass vendor validation. The AI-generated receipt has a plausible date, a realistic amount, and an authentic-looking layout. The only evidence it was never issued is in the pixel data.

“We reviewed three months of expense submissions after deploying Hesper. 4.2% of receipts showed pixel-level manipulation. Almost none had triggered any rule-based flag.”

— Head of Finance, European SaaS company (anonymised)

The architectural fix

The integration pattern for expense platforms is the same as for any document workflow: intercept the receipt before it reaches your OCR pipeline, analyze the raw image, receive a fraud score and findings in under 100 milliseconds, and route based on the result.

1. When a receipt is uploaded, before passing it to your OCR pipeline, send it to Hesper for analysis
2. Receive a fraud score (0–100) and an array of findings with pixel coordinates
3. If the score exceeds your threshold (typically 65–70 for expense workflows), route to manual review
4. Pass findings and coordinates to the reviewer so they can inspect the specific region
5. If clean, continue to your existing OCR and approval flow unchanged

What reviewers see

{
  "fraud_score": 87,
  "verdict": "LIKELY_FRAUD",
  "confidence": 0.94,
  "findings": [
    {
      "type": "pixel_manipulation",
      "description": "Clone stamp artifact in amount field — compression discontinuity",
      "region": { "x": 142, "y": 88, "w": 168, "h": 32 },
      "severity": "high"
    }
  ],
  "latency_ms": 78
}

The structured findings give reviewers an exact location to inspect. Instead of reviewing the entire receipt, the reviewer opens the document and checks the region at coordinates {x:142, y:88}. The review takes 15 seconds instead of 2 minutes. High-confidence cases (score >85, severity: high) can be auto-declined without review if your risk policy permits.

The business case

For expense platforms, the ROI calculation is straightforward. Take your monthly expense report volume, apply the industry average fraud rate (approximately 4–5% of reports contain manipulation), and multiply by the average fraudulent claim value. The savings from catching even a fraction of this exceed the cost of the detection layer by a significant margin.

For expense platforms building this detection into their product, the value proposition is different: pixel-level fraud detection becomes a competitive differentiator that justifies higher pricing with enterprise customers. The integration is one afternoon of engineering work, and the differentiation is permanent.

Frequently asked questions