When seeing is no longer believing: AI and document fraud
Businesses have spent years automating the processing of claims, applications and supporting documents. Generative AI now makes those documents far easier to fake, and one automation specialist argues that checking whether evidence can be trusted has to become part of the workflow itself.
Director,
Netcall Europe
Pol Brouckaert is Director of Netcall Europe. In this contributed piece he sets out his own view of how organisations should build fraud detection and validation into automated document processing, rather than treating it as a separate check afterwards. The views are his own.
For years, businesses have worked towards removing friction for customers and manual tasks for staff. If a customer submits a claim, application or supporting document, technology can classify it, extract the relevant information, validate it and pass that data into the systems that need it. What once required somebody to open a document and re-key information now happens automatically. Year after year, an increasing percentage of cases are automated end to end without human intervention.
But generative AI has introduced a complication. The documents and images entering those processes are becoming easier to manipulate. According to Inscribe's 2026 State of Document Fraud report, around one in 16 documents analysed showed signs of manipulation, fabrication or misrepresentation, while monthly volumes of detected AI-generated document fraud increased nearly fivefold between May and December 2025.
An insurance claimant could alter a photograph of vehicle damage. An employee might submit an edited receipt as part of an expense claim. Documents used to demonstrate income could potentially be changed before entering a mortgage application.
This creates a challenge for any organisation pursuing greater automation. It is no longer enough for a system to understand what a document says. Increasingly, it also needs to establish whether that information can be trusted.
Fraud is bigger than insurance
Insurance is perhaps the most obvious example. Claims processes depend heavily on evidence, from photographs of damage to invoices, medical documents and accident forms, but the same issue applies across many industries.
Mortgage providers rely on documents that demonstrate income and financial circumstances. Know Your Customer (KYC) and Know Your Business (KYB) processes depend on evidence of identity and company information. Public sector bodies may require proof when assessing benefits, issuing licences or granting permits. Logistics and customs processes also rely on certificates and official documentation.
The common thread is not the industry. It is proof. Whenever an organisation asks someone to provide evidence of their identity, address, income, expenses incurred or damages suffered, there is a question of authenticity.
The wider cost of fraud shows what is at stake. While not specific to document fraud, Nasdaq Verafin's 2026 Global Financial Crime Report estimates that fraud scams and bank fraud schemes caused $579.4 billion in losses worldwide in 2025.
That makes identifying potential fraud as early as possible increasingly valuable. Rather than treating fraud detection as a separate exercise at the end of an automated process, organisations need to think about how trust and validation can be built into the workflow itself.
Reading the document is only half the job
Historically, document automation was constrained by what technology could reliably read. A neat, typed form was one thing. Handwriting, checkboxes, drawings, unusual layouts and photographs were considerably harder.
Advances in intelligent document processing (IDP) have changed this significantly. Take motor insurance. The European accident statement can contain handwritten information, ticked boxes, signatures and even drawings showing what happened. Add photographs of the vehicles and supporting invoices, and the information required to assess a claim is spread across very different formats.
Modern IDP can increasingly interpret these less structured inputs, from submitted forms and supporting documents to images. That creates significant opportunities to automate processes that previously required considerable human intervention.
Yet greater capability creates greater responsibility. If an automated system is going to use a photograph or supporting document to help determine what happens to a claim, understanding its contents is only part of the task.
The integrity of that evidence matters too. Checks can look for signals such as inconsistencies or anomalies that indicate something may require closer scrutiny, alongside validating whether the information in a document makes sense. Currently, many organisations focus on sampling after the event, and some may only intervene if one client submits suspiciously more claims than expected. Modern document processing, however, can contribute to fraud detection at the point information enters a workflow, rather than relying solely on checks further down the line.
The important point is not that technology should make an absolute judgement that something is fraudulent. It is that the result of those checks can inform what happens next.
Automation needs to know when to stop
This is where the conversation around AI accuracy can sometimes become misleading.
Organisations understandably want AI systems to be accurate, particularly in regulated industries. But focusing on whether a system achieves an abstract accuracy score can distract from the more useful question: how much of the process can we safely automate?
Not every document needs to be treated in the same way. An organisation might set confidence thresholds for different stages of a process. Where the information is clear and the relevant validation checks are satisfied, a case can continue automatically. Where confidence falls below an agreed threshold, information is missing or something appears suspicious, it can be passed to a person for review or further fraud checks.
This is an important distinction. The goal should not be automation at any cost.
People still have a vital role, particularly when cases are complex or uncertain, or when empathy is a must. The opportunity is to use their judgement where it adds most value, rather than asking them to inspect every routine document by hand because the organisation has no way to tell the straightforward cases from the questionable ones.
For CIOs, CTOs and operations leaders, that changes the question from “Can AI process this document?” to “Under what circumstances are we comfortable allowing the process to continue automatically?”
Trust needs to become part of the workflow
This matters because extracting data is rarely the outcome a customer actually wants. Behind every process is someone waiting. Someone submitting an insurance claim is waiting for a decision. A person applying for a mortgage is waiting for an answer. A business completing a regulatory process wants to know whether it can proceed. All of them want certainty, and the document is simply one part of that journey.
As IDP becomes connected to wider automation, information can be classified and extracted before business rules and validation checks are applied. Straightforward cases can progress, while exceptions are directed to the people equipped to investigate them. That creates the possibility of faster services without pretending that every decision can, or should, be left entirely to AI.
It also changes how we should think about document fraud. The answer is not simply another fraud detection tool sitting alongside the process. Validation needs to become part of how information moves through the organisation.
Generative AI has made creating and manipulating digital evidence easier. Businesses cannot respond by abandoning automation and manually checking everything that comes through the door.
Instead, the next phase of automation needs to combine speed with appropriate scrutiny. Because when seeing is no longer enough to guarantee believing, trust has to become part of the process itself.