The Insurance Fraud Divide
Why some carriers are pulling ahead on fraud, and what the rest of the industry is missing
Property and casualty insurance fraud in the United States now costs the industry between $90 and $122 billion a year1. That figure has nearly quadrupled since the Coalition Against Insurance Fraud’s baseline estimate of $80 billion across all of P&C in 19952. The growth is happening at the same moment carriers are spending more on fraud analytics, anomaly detection, and AI-powered claims triage than at any point in the industry’s history. The two trend lines have been moving in the wrong direction, together, for nearly a decade.
Inside the industry, an approach is starting to come into focus.
A small but growing group of carriers, large data and analytics providers, and a handful of forward-leaning specialty insurers have been quietly outperforming their peers on fraud detection. They are catching more rings, paying fewer fraudulent claims, and defending more confidently when regulators ask how their AI made the decisions it did.
They aren’t winning because they bought better fraud models. They bought roughly the same fraud models everyone else bought. They are winning because of what sits underneath.
The dividing line in insurance fraud right now is identity. Specifically, some carriers have built a continuous and accurate view of who is who across all their operational data—from book to claims to providers to brokers, including all the external data they consume. The other carriers are running increasingly sophisticated AI on bad data. And the gap between these two camps is widening every quarter.
What exactly are the carriers in the winners’ camp doing that the rest of the industry has not yet caught up to?
Why More AI Has Not Meant Less Fraud
The intuitive answer when fraud losses keep climbing is to invest in more sophisticated detection. The industry has done exactly that, with mixed results.
Most carriers now run multiple anomaly detection models across claims, automated FNOL (First Notice of Loss) scoring at intake, network analysis tools in their SIU (Special Investigation Unit) functions, and increasingly some form of generative AI for document review.
The technical sophistication is real. The investment is real. The reduction in fraud is not.
The reason is that fraud detection is, underneath, an identity resolution problem dressed up as a fraud detection problem. A fraud ring is not detectable in any single claim. Detection can happen only through examination of the relationships between claims, claimants, providers, vehicles, addresses, and devices. If a claimant appears in your systems as Robert J. Henderson on one policy, Bob Henderson on a claim, and R.J. Henderson in a contributory database, your fraud model sees three different people. If a body shop’s name is spelled three slightly different ways across your medical billing fields, your network analysis sees three different providers. The model cannot detect the ring because the underlying data does not represent the ring as a ring. It represents it as a collection of independent, mostly unremarkable claims.
Gartner has been blunt about this dynamic. In a 2025 research note3, the firm observed that across systems pulling data from diverse sources for AI and analytics, the identifiers for the same entity tend to be inconsistent between data silos, and records describing the same person from different sources rarely match. The result, in Gartner’s language, is that AI and analytics systems cannot meaningfully join the data. In insurance terms, that means the fraud model cannot see what the SIU investigator could see if she had three weeks to manually reconcile records across systems. She does not have three weeks. The model has four seconds.
Carriers that have closed this gap have done so by treating identity intelligence as a foundational capability, not a feature inside any single application. They have built a layer that resolves entities continuously across messy, multi-source data, in real time, and they have wired the rest of their stack to use it. That is the comprehensive, essential work most of the industry has not yet done.
Five Practices to Meet the Moment
Across the conversations I have had with carriers, data providers, and SIU teams over the past two years, five practices consistently separate the carriers winning the fraud fight from the ones falling behind. None are theoretical. All five are operational at multiple carriers and at least one industry-scale data provider today. Here is what they are doing to meet the moment and win:
They resolve identity at FNOL, not at adjudication.
Most carriers do their identity work at the end of the claims process, if at all. A flagged claim goes to SIU. The investigators pull records, query external databases, reconcile inconsistencies, and eventually surface a connection. By that point the claim is months old, the payment may already be in motion, and the broader ring has had time to file additional claims at the same carrier or at others. The risks – and costs – are cascading.
On the other hand, the up-to-speed carriers have moved their identity work to the front of the process. When an FNOL arrives, the claim is enriched in milliseconds against a continuously updated, resolved view of every entity already in the carrier’s data: claimants, policyholders, providers, body shops, attorneys, addresses, phone numbers, vehicles. By the time the claim reaches the adjuster’s screen, the system already knows whether this claimant shares an address with a previously off-boarded policyholder, whether this body shop has appeared in suspicious billing patterns across the carrier’s book, and whether this vehicle has been the subject of prior claims. The decision the adjuster or the model makes is made on current, accurate, resolved data, not on a simplistic string match.
They treat external data as first-class, not as enrichment.
A common pattern in less mature programs is to treat external data sources as something you query when a claim is already suspicious. The carriers pulling ahead treat external feeds the way they treat internal data: continuously resolved into the same entity graph, available at the same speed, governed by the same access controls. ISO ClaimSearch contributions, National Insurance Crime Bureau data, sanctions and watchlists, business registries, public records, contributory databases from peer carriers: these are not enrichment sources for the leading carriers. These are the substrate on which today’s fraud model reasons.
The practical difference is enormous. A claimant whose third claim at your carrier looks unremarkable might be the same claimant whose seventh claim at another contributing carrier is clearly an illicit ring participant. If your identity layer resolves across both, you see the ring. If it doesn’t, you pay out on a fraudulent claim.
They expect messy data, and design for it.
Master data management, the dominant data discipline of the past two decades, was built around an instinct to clean. Eliminate variation. Converge on a single golden record. That instinct works for known internal entities in bounded systems. It fails for fraud detection, where the variation is often the signal.
Misspellings, aliases, prior addresses, differing dates of birth, transliterated names, missing digits, off-by-one phone numbers: in an MDM worldview these are noise to be cleaned away. In an identity intelligence worldview they are often the only clue that a ring is operating, that a synthetic identity is in the book, or that a relationship exists between two parties that nobody intended you to see. The variability is the clue. The carriers pulling ahead have internalized this and have built systems that resolve identity across messy data rather than rejecting it.
They make their identity layer explainable to humans, regulators, and auditors.
Black-box matching cannot survive contact with a regulator. The NAIC model bulletin on AI in insurance, now adopted by over half of US states4, places direct expectations on carriers to govern the data underpinning AI used in claims, underwriting, and fraud detection.
As the new threat horizon is closing in at velocity, regulatory expectations and standards are also accelerating. New York DFS Circular Letter No. 7 goes further on data integrity in underwriting5. In the UK and EU, FCA Consumer Duty and the new AMLR/AMLD6 package, which enters full force in 2027, all raise the bar on ongoing customer due diligence for any insurer in the AML/CFT perimeter6.
Across all these regimes, regulators now implicitly assume the carrier has, or can build, an identity layer it can explain. When the system says two records are the same person, or when it says they are not, a human should be able to see why. When the system says a relationship exists between two parties, the audit trail must show what evidence the system used and how it weighted that evidence. The carriers pulling ahead have built identity layers that can pass that test. Carriers that can’t keep up will increasingly find themselves explaining AI decisions they did not design and cannot defend.
They treat identity not as a fraud tool, but as governance infrastructure.
Probably the most important and least appreciated practice. The carriers winning the fraud fight do not think of identity intelligence as simply an add-on detection feature. They recognize this is foundational data infrastructure, governed deliberately and used by multiple downstream functions: underwriting, fraud, customer experience, marketing, compliance, regulatory reporting, and increasingly the AI agents that are starting to sit inside all of these functions.
The practical consequence is that the investment justifies itself across multiple budget lines, the governance is owned by a senior leader rather than dispersed across IT and SIU, and the capability gets continuously fed by the carrier’s broader data work—rather than starved by it. And rather than accepting cascading risk, they are rewarded with compounding value.
What This Looks Like at Industry Scale
One of the clearest published examples of a company with comprehensive identity intelligence is Verisk. Through years of acquisitions and organic growth, Verisk had accumulated different entity resolution teams, technologies, and processes across dozens of business groups inside the company. Standardizing identity resolution across the enterprise became a strategic priority for data quality, efficiency, and the trust of the carrier customers who rely on Verisk’s analytics. Verisk’s platform now resolves more than 1.6 billion records into over 420 million unique identities, with sub-second response and a sharp reduction in false positives7. The outputs of that platform are embedded inside the products that thousands of carriers across the industry consume every day.
What Verisk did at industry scale, individual carriers can do for their own books.
The pattern is the same: a foundational entity resolution layer, fed continuously by internal and external data, exposed to every downstream system that needs to reason about identity. The technology to do this exists now. The question is no longer whether it can be built. It is whether a given carrier will build it before the AI agents now sitting inside its claims, underwriting, and SIU workflows make another year of decisions on identity data that does not know who is who.
The Stakes Are Compounding, Not Linear
There is an argument to be made that the gap between the two camps is temporary. That the rest of the industry will catch up. That fraud losses will eventually moderate as carrier sophistication increases across the board.
What I am observing, though, is that the gap is compounding, not closing. Three dynamics drive this. First, the carriers that resolved identity early are now training AI agents on resolved data, which makes those agents materially better than agents trained on fragmented data.
Second, the carriers in the leading camp now exchange contributory data more efficiently with each other through resolved-identity layers, which means a fraud pattern surfacing at one carrier propagates to peers within hours, not quarters.
Third, the carriers that started this work three or four years ago are now writing the regulatory expectations the rest of the industry will be measured against, because the model bulletins and circular letters reference the practices they pioneered.
The economic stakes are real and they show up in the lines of business under the most stress. Combined ratios in commercial auto sat above 107 in 20248. Commercial multi-peril liability has not posted a combined ratio under 100 since 20158. In segments like these, the line between a viable book and a structurally unprofitable one is thin enough that the difference between fraud caught at FNOL and fraud paid at FNOL is, for some books, the difference itself.
Carriers that close the identity gap now will see it show up in their loss ratios over the next three to five years. Carriers that do not will see the inverse.
The Question Every Carrier Should Be Asking
The question is not whether to invest more in AI. Every carrier is doing that. The question is whether the identity layer underneath that AI is good enough to make the AI’s decisions defensible, the SIU team’s investigations productive, and the underwriting model’s risk pricing accurate. For most carriers today, the honest assessment is that, right now, they don’t know how far they are from good enough.
The carriers pulling ahead in fraud are not the ones with the most sophisticated models. They are the ones who decided, three or four or five years ago, that the data layer underneath the models was a strategic capability worth building deliberately. They are reaping the compounding benefits now.
The rest of the industry will spend the next three years catching up, and some of them will not make it.
The fraud divide is not a future problem. It is the live, competitive battlefront, and the losses are compounding as you read this.
Sources
- 1
Coalition Against Insurance Fraud, The Impact of Insurance Fraud on the U.S. Economy (2022 update), in partnership with Colorado State University Global White Collar Crime Task Force. P&C estimate of $90 to $122 billion is the property and casualty subset within the $308.6 billion all-lines total. Available at insurancefraud.org. - 2
Coalition Against Insurance Fraud, 1995 baseline estimate of approximately $80 billion in P&C insurance fraud, as referenced in the CAIF 2022 update report cited above. The 2022 study addresses all lines of insurance; the historical comparison is directional rather than line-equivalent. - 3
Gartner, Doing “Just Enough” Master Data Management for Analytics and AI, Lyn Robison, 14 October 2025. The paraphrase in this article reflects the published research note’s observation on identifier inconsistency across data sources. - 4
NAIC Big Data and Artificial Intelligence Working Group, tracking of state adoption of the Model Bulletin on the Use of Artificial Intelligence Systems by Insurers (originally adopted by NAIC in December 2023). As of late 2025, over half of US states have adopted the bulletin or substantially similar guidance, with twenty-four states formally adopting and additional states issuing related regulations. Current tracker available at content.naic.org. - 5
New York Department of Financial Services, Insurance Circular Letter No. 7 (2024), addressing the use of artificial intelligence systems and external consumer data and information sources in insurance underwriting and pricing. - 6
European Union Anti-Money Laundering Regulation (Regulation (EU) 2024/1624) and Sixth Anti-Money Laundering Directive (Directive (EU) 2024/1640), entering full force on 10 July 2027. Coverage extends to life insurance, certain non-life insurance lines, and brokers within the AML/CFT obligated-entity perimeter. - 7
Senzing, Verisk Deploys Enterprise-Wide Senzing Entity Resolution, customer case study, available at senzing.com/verisk-entity-resolution-senzing/. The 1.6 billion records and 420 million unique identities figures are drawn directly from this published case study. - 8
S&P Global Market Intelligence and AM Best, 2024 U.S. property and casualty industry segment results. Commercial auto combined ratio of 107.2 and commercial multi-peril liability combined ratio of 114.9 are from S&P Global Market Intelligence as reported May 2025. The overall P&C industry combined ratio was 96.5 in 2024, the strongest underwriting performance since 2013; the commercial liability lines remain materially above 100 despite the industry-level improvement.
