Document fraud in insurance: AI's new invisible front

Document fraud is nothing new in insurance. Fake supporting documents, altered invoices, doctored IDs and fraudulent bank details have long been part of the risks insurers learn to control.

What has changed is the scale. With generative AI now in everyone's hands, producing or altering a credible document has become easier, faster and harder to detect. A claim photo can be retouched in seconds. An invoice can be modified without leaving any obvious trace. A document can look perfectly consistent on its own, yet reveal an anomaly once cross-checked against the rest of the file.

For insurers, brokers and MGAs, document fraud is becoming a new operational front: invisible, diffuse, and potentially very costly. The question is no longer just whether to check documents, but when, with which signals, and at what scale.

Document fraud is going industrial

The real shift is accessibility. Not long ago, producing a fake invoice or altering a supporting document took time, skills or access to specialised networks. Today, widely available AI tools can generate or modify documents, images and evidence in minutes.

The cost of producing a fake document has collapsed while its quality has soared. Fraud is moving from a cottage industry to an industrial one.

The document becomes the critical entry point

In insurance, almost every process starts with, or is justified by, a document: ID, bank details, company registration, invoice, quote, photo, accident report, expert assessment, medical certificate, and so on.

This matters at two key moments:

• at underwriting: identity verification, supporting documents, compliance, consistency of the declared profile;
• at claim: evidence, invoices, quotes, photos, medical documents, reports, proof of loss.

The problem: if the check comes too late, the fraud is already inside the system. This is especially true at FNOL, where the signals are still fresh but rarely exploited in traditional workflows. And the faster and more digital the journey, the more those checks need to run in real time, inside the journey itself.

AI is also changing the nature of fraud

Document fraud is no longer just about checking whether a PDF has been edited. It is becoming multimodal and contextual.

Until now, defrauding mostly meant tampering with an existing document: changing an amount, a date, a name. Those manipulations left traces, in the metadata or in the visual itself. With generative AI, a fake is no longer a modification, it is a creation. An invoice generated from scratch carries no trace of tampering, because nothing was tampered with.

And fraud no longer stops at a single document. A fraudster can now produce a complete, coherent file: the photo of the water damage, the contractor's quote, the matching invoice. Every piece is clean. It is the whole that is fake.

Detection now needs to cover:

• metadata anomalies;
• inconsistencies in dates, amounts, names or locations;
• AI-generated or AI-retouched content;
• documents that are individually credible but inconsistent with each other;
• suspicious patterns across an entire file.

That is what makes the problem harder: a document can look valid in isolation, then turn suspicious once cross-checked against the other pieces. The unit of analysis is no longer the document, it is the file.

An AI versus AI race

The market is entering an arms race: AI makes fraud easier, but it is also becoming essential to detect it.

Verisk reports that 98% of surveyed insurers believe AI editing tools are fuelling digital fraud, while only 32% say they are very confident in their ability to detect deepfakes. Fraud involving deepfakes, synthetic voices and identity manipulation is also accelerating, going well beyond "simple" documents.

In France, L'Argus de l'assurance reports close to one billion euros of fraud detected in 2025, including identity theft, fake bank details and document fraud. And that is only the fraud that was detected.

Are manual checks becoming insufficient?

Manual review remains essential, but it can no longer be the only line of defence.

Historically, detection relied on the handler's eye: a pixelated logo, an inconsistent font, a surprising amount. That vigilance worked as long as fraud stayed artisanal and volumes stayed manageable. Both conditions have now collapsed at once: digital journeys multiply incoming documents, and AI-generated fakes no longer show the visual flaws that used to give fraudsters away.

Add to that the pressure on processing times. Customers expect fast answers at underwriting and at claim alike, and every additional manual check slows down the journey for every file, including the 95%-plus that are perfectly legitimate.

Its limits:

• too slow at high volumes;
• costly;
• dependent on each handler's experience;
• hard to keep consistent across teams;
• often triggered too late in the journey;
• poorly suited to invisible or cross-document anomalies.

The goal is therefore not to replace humans, but to surface the right signals to them at the right time.

From spot checks to embedded detection

Faced with this new reality, the answer cannot rest on manual or one-off checks alone. Handlers must remain the decision-makers, but they need tools that analyse documents the moment they are submitted, detect anomalies invisible to the naked eye, and cross-reference information across the whole file.

That is precisely the role of a document fraud detection module: turning every incoming document into an automatic checkpoint, without slowing down the journey.

At Korint, this approach is embedded directly into underwriting and claims workflows. It is not an analysis tool sitting next to the process: the module has access to all the data captured throughout the journey, so it can cross-check every supporting document against the declared information, not just analyse the document in isolation.

In practice, every document is automatically classified, analysed and checked: retouched photos, altered invoices, metadata anomalies, inconsistencies between pieces. But rather than producing a score per document, the module aggregates these signals into a fraud score at the level of the claim itself, where the decision is actually made.

And because detection lives inside the business workflows, it triggers actions: above a certain score, a verification task is created for the handler or an override is required. At underwriting, the detected risk level can even feed into pricing. Teams keep the final say, in combination with each insurer's business rules.

The goal is not to replace human expertise, but to give insurers better control over their risk, and to focus that expertise where it creates the most value: on the files that deserve it.