Insurance Fraud Statistics and the 1980s Estimate Behind Them

Open any article on insurance fraud and the first number is the same. Three hundred and eight point six billion dollars a year, every year, in the United States alone.

It appears in vendor marketing, in regulator blog posts, in conference keynotes. In May 2026 it appeared in a press release from SAS, alongside the claim that roughly one in ten property and casualty losses involves fraud.

Both numbers are worth knowing. Neither is a measurement.

This piece is about what the fraud statistics actually are, where they come from, and which ones you can take into a pricing or SIU conversation without being embarrassed.

The short version: the famous American number is an estimate assembled in 2022, and the most useful figures in the field are the smaller ones, from the markets that count detected fraud instead of modelling total fraud.

The number everyone opens with

The $308.6bn comes from The Impact of Insurance Fraud on the U.S. Economy, published by the Coalition Against Insurance Fraud on 26 August 2022. The data collection and analysis were done by the Colorado State University Global White Collar Crime Task Force under Dr Michael Skiba, and the report was peer-reviewed by the IASIU, the NICB, the APCIA and the Insurance Information Institute.

It was the first study of its kind in more than 25 years. The figure it replaced was $80bn, circulated by the Coalition since around 1995. Run that through CPI and you get $155bn, so a little over half of the increase is inflation catching up with a number nobody had revisited since the Clinton administration.

The study is careful about this. It presents itself as a set of line-by-line estimates, explains each formula, and asks readers with better data to come forward. That is a reasonable way to publish. What happens afterwards is the problem: the total gets quoted as though it were counted.

Where the 10% came from

Life insurance is the largest line at $74.7bn, just under a quarter of the total. The derivation is two steps. The Insurance Information Institute reported $747.4bn of US life insurance benefits and claims paid in 2020. The task force multiplied that by 10%.

The 10% is the interesting part. The report says the figure is “widely circulated” by the NICB, the III, the Coalition, the NAIC and the IASIU as the share of claims that are fraudulent. Follow it back one more step, to the III’s own account, and the origin is specific.

“In the late 1980s, the Insurance Information Institute interviewed claims adjusters and concluded that fraud accounted for about 10 percent of the property/casualty insurance industry’s incurred losses and loss adjustment expenses each year.”
Insurance Information Institute, quoted in the Coalition’s 2022 study

So the largest component of the number the industry quotes in 2026 is total life payouts multiplied by a rule of thumb derived from adjuster interviews conducted around forty years ago, in a different line of business.

That same 10% produces the $38bn property and casualty figure the report cross-checks against. It is also the one in ten that SAS repeated this year.

What the other lines rest on

Workers’ compensation is $34bn, split into $9bn of claimant fraud and $25bn of premium fraud.

The claimant half is written premium multiplied by 16%, a share of claims the task force had previously found to be suspicious. The premium half starts from a $3bn estimate for California and scales it by population: California is 12% of the United States, so multiply by 8.3.

Medicare and Medicaid at $68.7bn is the midpoint of an AARP figure of $60bn and a GAO figure of $77.4bn. Auto premium fraud at $35.1bn is the midpoint of $31.6bn and $38.7bn. Disability at $7.4bn is a 2019 improper-payments figure from the Social Security Administration, CPI-adjusted forward.

None of this is dishonest and none of it is hidden. It is all in the document. But an estimate built from one assumed percentage, one population scaling and three midpoints is an order-of-magnitude marker. Treating it as a measured loss cost is how you end up defending a business case you cannot support.

UK Insurance Fraud Statistics (2024)

The United Kingdom counts detected fraud and publishes it annually, which makes it the more useful dataset even for an American reader.

On 17 November 2025 the Association of British Insurers reported £1.16bn of fraudulent general insurance claims detected during 2024, up 2% on the £1.14bn found the year before. Insurers identified 98,400 fraudulent claims, up 12% from 81,100.

Motor was just over half of it: 51,700 claims worth £576m, up 5%. Property fraud came to £189m, up 11%. The most common single behaviour was claim exaggeration, up 10%.

“With insurers continuing to detect over £1bn worth of bogus claims, the fight must continue.”
Mark Allen, head of fraud and financial crime, Association of British Insurers

The average case is getting smaller

Put the two ABI figures against each other and something shows up that neither headline mentions. Claim count rose 12% while value rose 2%, which means the average detected case got cheaper.

On the ABI’s own totals that is roughly £14,000 per detected case in 2023 and about £11,800 in 2024, a fall of something like a sixth in a single year. The totals are rounded, so treat the exact figure with care, but the direction is not ambiguous.

Which points at what kind of fraud is growing. Not organised rings staging catastrophic injury claims, but volume: exaggerated repair costs, added items, an injury that lasts longer on paper than in life. That is a detection problem in the ordinary flow of claims rather than a special-investigations problem, and it lands on rules and referral thresholds rather than on investigators.

Application Fraud and Ghost Broking

The figure in the ABI release that gets least attention is the largest count in it. Insurers prevented 684,800 cases of application fraud in 2024, up 7.4%.

Seven times as many prevented applications as detected claims. Fraud is being stopped at the point of sale far more often than at the point of loss, which is where you would want it stopped, and it says the underwriting question set is doing more anti-fraud work than the claims file.

Ghost broking sits inside this. The Insurance Fraud Bureau recorded a 52% rise in ghost broking activity between 2022 and 2024: fake or manipulated policies sold on social media, usually to drivers aged 17 to 25, either fabricated outright or bought genuinely and then altered on address, occupation or named driver to cut the premium.

The victim discovers it at first claim, when there is no cover.

Synthetic evidence is the part nobody is ready for

A claims handler used to be able to assume a photograph was a photograph. That assumption is gone, and the industry knows it is gone without being ready for it.

A survey by the Association of Certified Fraud Examiners and SAS found that only 7% of anti-fraud professionals considered their organisation more than moderately prepared to detect or prevent AI-assisted fraud. Among the insurance respondents, not one reported more than moderate confidence.

“With just a few prompts, fraudsters can use generative AI tools to create, enhance or erase visual evidence to support a false insurance claim.”
Franklin Manchester, Principal Global Insurance Advisor, SAS

Erase is the word to sit with.

Adding damage that was never there is one attack.

Removing pre-existing damage from a photograph, so a claim that would have been declined for wear becomes payable, is quieter and harder to catch, because nothing in the image looks wrong.

The defence is not really image forensics.

It is the surrounding record: when the file was created, what device made it, whether the same address has appeared on four unrelated claims this month, whether this phone has been claimed on before. Zurich has described exactly that pattern work, flagging several claims filed at once for different people at one address, or repeat handset claims traced to a single buyer.

Which is a data problem before it is an AI problem, and it is the same data problem that shows up in insurance claims data generally.

What to actually do with these statistics

Three practical conclusions come out of the numbers above.

Stop using $308.6bn as a business case. It will not survive a finance director who reads the methodology, and there is no need for it. A carrier’s own detected-fraud rate, its own referral hit rate and its own average case value are all measurable in-house, and they are the figures a business case should rest on anyway.

Second, weight the effort towards volume. If the average detected case is getting smaller and application fraud outnumbers claims fraud seven to one, the return sits in rules that run on every submission rather than in investigator headcount. Referral criteria you can change when a pattern shifts are worth more than a model refresh cycle measured in quarters.

Third, keep the record. Every AI-era fraud defence described above depends on being able to reconstruct who submitted what, when, and what the system did with it. An insurance audit trail written as a by-product of running, rather than assembled afterwards, is what makes a declination defensible.

That is where the platform question comes in, and it is narrower than vendors make it sound. Openkoda is not a fraud analytics product and would lose a feature comparison against one.

What it does is keep the parts you need to change in configuration you own: referral rules and thresholds in an automated underwriting system a business user can edit, questions added to an application without a release, and the audit record generated by default. Fraud patterns move faster than a development calendar, so the ability to change the rule next week is worth more than the sophistication of the rule you shipped last year.

The honest state of the evidence

Insurance fraud is large, it is growing in count, and almost nobody knows its true size.

The best American estimate is four years old and built on assumptions its own authors documented and invited challenge to.

The best measured figures come from a market a fifteenth the size of the United States. And the newest attack surface, synthetic evidence, has no loss statistics at all yet, only a survey saying the people responsible do not feel ready.

None of that argues for doing less. It argues for measuring your own book instead of quoting someone else’s, and for building the kind of record that will let you count properly when somebody finally asks.