FraudWorks
Try the demo
NewStandard and Deep checks in the live demo

Know if a document is real.

Upload a receipt, invoice, bank statement, or pay stub. FraudWorks checks the file, the pixels, and the math, and shows you exactly what looks wrong.

Check a document freeSee how it works
Receipt with edited total, page 1
VerdictLikely tampered
Risk Score95/100
  • Total does not add upSubtotal 211.59 + tax 18.78 = 230.37, but the printed total is 260.37.
  • Text replaced by drawing over itA shape hides '230.37', and '260.37' is printed on top of it.
  • Printed number differs from the text layerThe page shows '260.37', but the PDF text says '230.37'.
  • Number in a different font'260.37' is set in Helvetica, which appears nowhere else on the page.

provenance · structure · content

In a 2026 study of more than 3,000 AI-forged receipts and forms, human reviewers spotted the fakes only about half the time. Source

What it catches

Built for the fakes your team actually sees

Four kinds of document fraud, each caught by more than one check, so no single weak signal decides.

AI-generated documents

Receipts and invoices made by image generators, caught from Content Credentials and metadata the generator leaves behind.

Edited totals and dates

A white box over the old number, a new number on top. Caught in the PDF structure, the fonts, and the math.

Forged bank statements

Balances that do not follow from the transactions, and statements re-saved in a consumer PDF editor.

Fake pay stubs

Net pay that does not add up, Social Security and Medicare at the wrong rates, year-to-date figures that go backwards.

The result

Every result shows its work

No black-box score. Each result lists what fired, why, and where on the page.

  1. 1Verdict. Four plain outcomes. A clean result says no issues were found, never that a document is genuine.
  2. 2Risk Score. 0 to 100. Several medium signals add up; one weak signal cannot dominate.
  3. 3Highlighted region. The exact spot on the page each signal points at.
  4. 4Signals. Each one with evidence, like the arithmetic that does not add up.
Likely tampered

Closing balance does not add up. Edited with a consumer PDF editor.

chk_A34Y8C2S1ATVS7P99Y54HAZZZV · engine 1.0.0

Risk Score80/100
Document typeBank statement
Created withSejda PDF Desktop
Edited after creationFlagged: Yes
AI provenanceNot found
Math checksFlagged: Failed
Pages1
Page 1 of the document

How it works

Four layers, not one model

AI-image classifiers alone are wrong too often. FraudWorks checks provenance, structure, pixels, and the math, then combines them.

01

Provenance

Reads Content Credentials, EXIF, XMP, and IPTC.

c2pa_ai_generated
02

File structure

Looks inside the PDF: producers, revisions, fonts, cover-ups.

pdf_text_overlay
03

Pixels

Compression history, noise, and cloned regions in photos and scans.

image_noise_inconsistency
04

Content logic

Totals, taxes, running balances, dates, and routing numbers, checked in code.

math_total_mismatch

Looks real. Isn't.

The edit a reviewer scrolls past

This receipt's total was covered with a white box and retyped. On screen it looks fine.

Receipt with edited total, page 1

Nothing obvious

The totals look right and the receipt looks clean. Switch to "What FraudWorks sees" to see the 5 signals behind it.

Use cases

Wherever a document decides a payout

Expense receipts

Catch fake and inflated receipts in expense reports.

Receipt with edited totallikely tampered · risk 95

Rental applications

Check pay stubs and bank statements applicants send.

Pay stub with wrong taxesneeds review · risk 48

Lending

Review statements and income documents in a loan file.

Bank statement with edited balancelikely tampered · risk 80

Rebates and claims

Stop fake purchase receipts from claiming payouts.

AI-generated receiptai generated · risk 93

Privacy

Sensitive documents, handled carefully

Statements and pay stubs carry names, account numbers, and income. How we handle them.

Demo uploads are never stored

They are analyzed in memory and discarded when the result is returned.

No training on your documents

We do not train models on documents you check.

Encrypted in transit

Every upload and result travels over TLS.

Developers

An API is on the way

Every check returns the same JSON the dashboard uses, so automating it later is a small step. The public API is not open yet.

Read the result format
GET /v1/documents/checks/{id}
{
  "id": "chk_VCV137RGZJKN4G30X4PHCTZZ1P",
  "verdict": "likely_tampered",
  "risk_score": 95,
  "signals": [
    {
      "id": "math_total_mismatch",
      "title": "Total does not add up",
      "page": 1
    },
    {
      "id": "pdf_text_overlay",
      "title": "Text replaced by drawing over it",
      "page": 1
    }
  ]
}

Check your first document in under a minute.

Try the live demo on a sample or on your own file. No account needed.

Try the live demoSee pricing