Learn

AI-generated invoice fraud: when the document was never real

TL;DR: Generative AI makes convincing fake invoices inexpensive to create, including plausible line items and a real vendor's visual identity. Ledger checks cannot identify a well-made fake when it creates a clean, internally consistent entry. Detection requires reading the document and testing its claims against delivery, purchase-order, vendor, and payment history.

What changed?

Invoice fraud once required more manual effort per attempt, and that effort often left visible errors: inconsistent layouts, reused templates, or incorrect tax calculations. Generative tools can now produce internally consistent documents at scale and imitate the format of a real supplier. They also make related tactics easier, such as credible bank-detail change requests or paperwork for a dormant vendor.

This is why ledger-based AP controls and document-level controls answer different questions. The review and approval model matters just as much as detection.

Why can't ledger-level controls see it?

The fraudster controls the fields those checks inspect. A fabricated invoice can pass a duplicate check because it is new, pass a format check because it is consistent, and appear to match if its numbers were selected for that purpose. The ledger entry can be clean even when the document behind it is false. The investigation therefore has to extend beyond the ledger.

What does document-level detection look for?

It looks for contradictions between the invoice and information the document cannot control:

  • No delivery behind it. No delivery note, goods receipt, or purchase-order history supports the goods being billed.
  • Vendor-master timing. Bank details changed shortly before a payment run, or a long-dormant vendor suddenly began billing.
  • Pattern breaks. Amounts sit just below approval thresholds, invoices are split, postings occur at unusual times, or numbering breaks from the vendor's history.
  • Document forensics. Metadata, fonts, or layout patterns differ from the vendor's genuine invoice history.

No single signal proves fraud. The strength of a case comes from the combination of signals and the evidence behind them.

What should finance teams do now?

First, review the past with a read-only lookback audit covering 12–24 months. Then run the same checks before payment, routing ambiguous cases to a human with the relevant evidence attached. Agents can investigate and assemble the case; finance retains the decision. The controlling and close use case shows how anomaly sweeps and evidence review fit into the control environment.