Insurance fraud rarely announces itself because most of the time, it looks like paperwork.
In 2023, Verisk analyzed a sample of 768,000 claim images and found nearly 2,000 photographs submitted more than once, recycled across 1,475 separate claims and representing roughly $5.3 million in fraudulent payments. Those images weren't caught during investigation or payment review, so they moved through intake on day one, unremarkable in a queue full of unremarkable documents. This isn't just a one-off problem; it's systemic. Fraudulent claims like these exploit the gaps in intake processes, proving that traditional approaches to fraud detection no longer suffice.
Fraud detection has historically lived at the back end of the claims lifecycle, in investigations, adjuster escalations, and payment audits, long after the documents that carried the signals were filed and forgotten. Without the tools to analyze intake documents at volume, downstream review was simply where detection was possible. That constraint is harder to justify today, when U.S. property claims volume rose 36% in 2024, driven in part by a 113% surge in catastrophe claims, and the fraud moving through the door is getting harder to spot, not easier. It's not just the volume that's daunting; it's the evolution of fraud tactics. Fraudsters are leveraging AI themselves, generating synthetic identities and hyper-realistic fake documentation that even seasoned adjusters find indistinguishable from legitimate submissions.
As claims volumes climb and fraud schemes grow more sophisticated, insurers must shift detection upstream to the first notice of loss (FNOL), where the signals arrive within the claim itself. Catching them requires AI that can analyze unstructured intake documents at scale, governed carefully enough to hold up under compliance scrutiny and transparent enough not to undermine the customer experience in the process.
The detection gap
Most insurance fraud detection systems intervene too late in the claims lifecycle, allowing suspicious activity to move through intake and processing before it's ever flagged. Today's fraud detection workflows were built around a referral model: A claim comes in, an adjuster reviews it, and if something looks wrong, the case gets escalated to a Special Investigation Unit (SIU) for formal investigation. By the time an investigator opens the file, the claim has already cleared intake, moved through early processing, and in many cases already shaped how much money the carrier has set aside to pay it.
The SIU model has two weaknesses, and today's fraud environment exploits both. The first is capacity. For carriers processing thousands of claims a day, it is unrealistic to examine even a small fraction of suspicious claims through traditional investigative channels as SIUs were built to investigate, not to screen. SIUs can't handle today's volume or complexity. Investigators are drowning in cases, and the backlog means legitimate claims may wait longer for resolution. Second, the SIU referral model depends on fraud being visible enough for an adjuster to catch it. Fraud today is designed to avoid suspicion – to look legitimate. By the time it reaches an SIU, the damage is often already done. The system is reactive when it needs to be proactive.
Twenty to thirty percent of insurance claims now include altered images, fabricated documents, or synthetic medical reports — materials that arrive looking legitimate and give adjusters no reason to look twice. Consider a fabricated repair estimate submitted alongside doctored photos of vehicle damage. On the surface, everything appears consistent — until you realize the same images were submitted to three other carriers.
The National Insurance Crime Bureau projected identity theft-related insurance fraud to rise 49% by the end of 2025, with nearly a quarter of those cases involving synthetically generated identities. A growing share of what arrives at intake is indistinguishable from a legitimate submission.
A fraud detection model that waits for something to look wrong is never going to catch fraud designed to look right.
Fraud signals appear at FNOL
Fraud signals usually come within the claim itself. FNOL is the moment insurers receive raw claim evidence for the first time –photos, repair estimates, invoices, police reports — before the file has been touched by multiple reviewers or supplemented with additional documentation. Everest Group found that over 60% of manually filled FNOL forms contain errors or unreadable data, leading to delays and rework. Most of the evidence arriving at intake is unstructured, meaning the documents come in all sorts of formats that traditional claims systems simply weren't designed to handle at scale. If a policyholder adds a made-up detail deep in a PDF, there's no neat data for systems to flag an anomaly.
That's where technology can change the calculus. Take Aviva, a global insurance leader serving more than 18 million customers across the U.K., Ireland, and Canada. Like most carriers, Aviva was managing growing fraud volumes against investigation workflows that manual processes couldn't keep up with. Reviews were slow, operational risk was climbing, and fraud moving through intake had no mechanism to catch it at the source. After automating fraud detection and investigation workflows, Aviva uncovered more than 11,000 fraudulent claims worth £116 million, increased fraud detection rates 39% year over year, and now stops an average of 30 bogus claims every day by automating detection at intake. Aviva's success story demonstrates how transformative AI-powered fraud detection can be. By shifting detection upstream to FNOL, they not only improved their fraud detection rates but also reduced operational costs and improved customer trust — showing that fraud prevention and customer satisfaction can go hand in hand.
Applied to unstructured intake documents, technology can classify, extract, and cross-reference content at a speed and scale that manual review cannot match, surfacing anomalies before a claim moves further into the workflow. A wrong call at intake can delay a legitimate claim or produce a denial that regulators and policyholders will challenge. This is where AI thrives. Unlike manual review, which relies on a human eye catching anomalies, AI can identify patterns across thousands of claims, flagging inconsistencies in real-time. It's not just about speed — it's about precision, transparency, and scalability.
The clock now starts at FNOL
Deloitte projects P&C insurers could save between $80 and $160 billion by 2032 through AI-powered detection deployed across the claims lifecycle. However, that projection assumes carriers deploy it where the exposure actually starts. A detection model still built around SIU referrals, still dependent on fraud announcing itself through an adjuster's review and blind to what's arriving in unstructured documents at intake, isn't positioned to capture those savings.
The schemes coming through the door today were designed around exactly that model. Catching them requires looking at the evidence when it first arrives, with systems built to read what manual review can't, and governance rigorous enough to make those calls defensible. The stakes couldn't be higher. Fraud prevention is no longer just about saving money — it's about maintaining trust in the system. Policyholders expect their claims to be handled efficiently and fairly, and regulators demand compliance at every step. Getting it wrong at FNOL isn't just a missed opportunity; it's a reputational risk no carrier can afford.
(Photo credit: MYKHAILO KUSHEI/ Adobe Stock)
Nick Hoppenjans serves as Industry Consultant at Tungsten Automation.
© Arc, All Rights Reserved. Request academic re-use from www.copyright.com. All other uses, submit a request to TMSalesOperations@arc-network.com. For more information visit Asset & Logo Licensing.