The rise of AI-generated invoice fraud

Over the past 18 months, we have seen a measurable shift in the quality of fraudulent documents submitted through our customers' pipelines. Documents that would have required professional graphic design skills to fake two years ago are now being produced by non-technical fraudsters using freely available AI tools. The volume has also increased: AI-generated document fraud is up 400% since January 2024.

Invoice fraud is the most common entry point. Invoices are the highest-volume document type in most accounts payable and expense workflows, which makes them the highest-value target. A convincing fake invoice is now one of the most accessible fraud vectors in corporate finance.

What changed

The arrival of capable AI image editing models — and more recently, document-specific generation models — has fundamentally changed the cost structure of invoice fraud. Previously, creating a convincing fake required either access to the original digital file, graphic design skills, or a willingness to pay someone who had both. Today, a fraudster with a smartphone and a legitimate invoice as a reference can produce a convincing fake in under 10 minutes.

The specific capabilities that matter are: inpainting (editing a specific region of an image while preserving the surrounding context), text rendering (generating text that matches the font, weight, and spacing of the original), and metadata reconstruction (producing a file with plausible creation and modification timestamps).

Why this matters for finance teams

The fraud detection stack at most companies was built for an older threat model. Rule-based systems — flag duplicates, check amounts against policy limits, validate vendor names — catch opportunistic fraud. They fail against AI-generated documents because the document is indistinguishable from a legitimate one at the text level.

This creates an asymmetry: the fraudster's cost to produce a convincing fake has dropped to zero, while the defender's detection rate using traditional tools remains at 22–41%. The expected value of fraud has increased dramatically. This is reflected in the volume: we see more attempts, not fewer, as the economic calculus has shifted.

“The fraudster can generate a document that complies with all your rules. The fraud is only detectable at the pixel level — which is a level that no OCR-based system reaches.”

— Hesper AI Threat Research team

What pre-OCR detection does differently

Hesper AI operates before your OCR pipeline reads the document. We analyze the raw pixel data — detecting generation artifacts, editing artifacts, compression inconsistencies, and font rendering anomalies that are the byproduct of AI-based document manipulation.

• Generation artifacts left by diffusion models in synthetically created documents — subtle statistical patterns in pixel distributions that real documents do not exhibit
• Editing artifacts from inpainting and clone stamp operations — compression discontinuities at region boundaries
• Compression inconsistencies in regions that have been modified — JPEG re-encoding creates characteristic artefact patterns
• Font rendering anomalies that indicate character replacement — sub-pixel rendering differences between original and inserted text

The output is a fraud score from 0 to 100, a verdict (LIKELY_FRAUD / INCONCLUSIVE / LIKELY_GENUINE), and an array of findings with pixel coordinates, severity ratings, and human-readable descriptions. The structured output routes directly into your existing review queue.

What to do about it

1. Before passing each invoice to your OCR pipeline, send the raw image to the Hesper API
2. Receive fraud score (0–100), verdict, and findings array in <100ms
3. If score exceeds your threshold (typically 65–75 for invoice workflows), route to focused review queue
4. Pass the findings and coordinates to reviewers so they know exactly where to look
5. If clean, continue to your existing OCR and approval flow unchanged

Frequently asked questions