Insurance Claims Fraud Detection: Totals Rise, but AI’s Share Remains Unproven

The use case

In the first half of 2024, Allianz UK reported that a tool called Incognito, described as "a supervised machine learning tool that flags suspicious claims for expert investigation", produced £37.7m in fraud savings and a 10% rise in its detection rate (operator-reported, first half of 2024, republished by a vendor)⁵. The figures show what these systems are for. They rank incoming claims so that a limited number of investigators look first at those most likely to be fraudulent. A supervised model is trained on past claims that investigators labelled as fraud or not fraud, and then scores new ones.

Scope. Here, AI means machine learning that scores claims for fraud, including unsupervised methods such as isolation forests, which look for claims that differ from the rest¹¹. Several older technologies and human services do similar work and are not credited to AI in this paper. One Asian insurer replaced "a business-rules system plus manual handler review" (vendor-reported, date not shown)⁸. In the United Kingdom a police unit, IFED, reports detecting about 230 fraudulent claims a day (reported by the ABI, 2024)². Application fraud, underwriting models and health-claims pricing are out of scope. Fraud committed with AI, such as claims supported by AI-generated images, is noted by Aviva but is not covered here⁷.

Where it is running

Industry totals. The Association of British Insurers (ABI) reported £1.1bn of fraudulent general insurance claims detected in 2023, across 84,400 cases (industry-wide, insurer-reported, all methods)². For 2024 it reported £1.16bn across more than 98,400 claims, with cases up 12% (industry-wide, all methods, as reported by Express & Star)³. Neither page names a detection method. Detected totals rose in both of the latest industry releases read, but none of the pages separates better detection from more fraud, or machine learning from every other method.

United States. The Coalition Against Insurance Fraud estimated the annual cost of insurance fraud at $308.6bn (August 2022, a compiled estimate built mostly on 2016-2020 data, which the authors call "impossible to measure accurately")⁴. For property and casualty claims the estimate assumes a 10% fraud rate⁴. It is a cost estimate, not a measure of detection or of AI, and it is more than three years old. No newer figure was found.

Geography and older deployments. The evidence is dominated by United Kingdom operator reports. Deployments in Asia come only from vendor material. An earlier Indonesian deployment, at PT Asuransi Tokio Marine Indonesia, was announced in September 2020 with no results published¹⁰.

What the evidence shows

No source read reports a randomised trial, an A/B test or a control group for an insurer's fraud model. The strongest evidence is therefore retrospective: models tested on past claims.

Peer-reviewed studies. A study of one Spanish insurer's property claims found 2,641 confirmed frauds among 303,166 claims, a base rate stated as 0.8% (retrospective, one insurer, no control group, data from 2015-2016)¹. The author reports a precision of 88% and a recall of 96% for later-period tests, while the cross-validation results in the same paper show precision of about 0.67 to 0.69¹. Precision is the share of flagged claims that prove fraudulent. When fraud is below 1% of claims, a small loss of precision means many legitimate claims sent for investigation. Precision here is also measured against the insurer's own fraud labels, so it shows agreement with past investigator judgement, not an independent test of what is fraud.

A peer-reviewed study of three Saudi healthcare providers reported random forest accuracy of 98.21% on 396 records after synthetic balancing (retrospective, claims from January to May 2022, no control group)¹². The authors list "a small sample size, a limited number of features or variables" as limits¹². Because the data were artificially balanced, the figure is not a real-world accuracy. A 2023 article in the Journal of Risk and Insurance reports a field experiment with an insurer and concludes that supervised and unsupervised methods are "complements rather than substitutes" (abstract only; no performance figures were read)¹¹.

Independent review. A systematic review of 33 studies from 2016-2025, covering health insurance only, points to "severe class imbalance" and "misleadingly high accuracy"¹³. It reports that "relatively few studies investigated model interpretability, fairness, transparency, or regulatory compliance" and calls for "prospective validation using real-world health insurance data" (peer-reviewed review, published October 2026)¹³. Its conclusion is that AI shows "considerable potential"¹³.

Operator and vendor results. Allianz reports £37.7m in savings and a 10% rise in detection from Incognito (operator-reported, first half of 2024)⁵. The source does not say whether savings are gross or net of investigation cost, or how the 10% was measured. A vendor case study of a Pan-Asian insurer reports savings of over $100k in the first months and says 80% of flagged suspicious cases "would have gone undetected before" (vendor-reported, date not shown)⁸. That figure counts cases newly found, not cases correctly flagged, and no precision was given⁸.

What it takes: cost, integration, governance

Cost. No source read publishes a price or cost per claim for fraud-detection software.

Data and integration. The Pan-Asian insurer consolidated data from three legacy systems and added third-party data, and flags claims in real time before payment (vendor-reported, date not shown)⁸. At Tokio Marine the system uses policy data, claims data, estimates and damage photos, and alert handling was adjusted to claim handlers' workflow; the sequence from proof of concept about 2019 to launch in 2022 indicates a multi-year effort (vendor-reported, date not shown)⁹.

People. Where workflows are described, people make the final call. Incognito flags claims "for expert investigation"⁵, and Tokio Marine handlers keep responsibility for final judgements and explanations⁹. Lemonade said in 2021 that claims flagged by facial recognition for appearing under different identities go to human investigators (operator-reported)¹⁵.

Rules. In the United States, the National Association of Insurance Commissioners (NAIC) adopted a model bulletin on insurers' use of AI on 4 December 2023, and it expressly covers fraud detection¹⁶. As of 8 October 2026, 29 jurisdictions (27 states, the District of Columbia and Puerto Rico) had adopted it¹⁷. Colorado's amended Regulation 10-1-1 took effect on 15 October 2025 for private passenger auto and health insurance¹⁸. In the European Union, recital 58 of the AI Act states that AI used to detect fraud in the offering of financial services should not be considered high-risk, while life and health insurance risk assessment and pricing are treated as high-risk¹⁹. The wording of the Annex that lists high-risk uses was not confirmed from an official text. The European insurance supervisor EIOPA issued an opinion on 6 August 2025 covering AI that is not high-risk, including fraud detection, and says it sets no new requirements²⁰. In the United Kingdom, draft guidance from the data protection regulator, updated on 31 March 2026, applies new rules only to decisions that are solely automated and significant²¹.

Where it falls short

The test of these claims is how they were designed. No control group was found for any insurer result, and every performance figure comes from one insurer or one dataset. Reported totals are gross: no source gives savings net of investigation cost, or counts legitimate claims delayed by a false flag. Operator totals such as Aviva's £233m (2025, all methods) are not attributed to AI⁷, and Allianz's £92.6m credits machine learning alongside voice analytics without a split (first half of 2025)⁶.

Whether gains persist and reach those who need them. No source read reports sustained performance over several years or tests of how the model behaves for different groups. The independent review notes that few studies examine fairness¹³.

Legal challenge. In Huskey v. State Farm, filed on 14 December 2022, plaintiffs allege that State Farm's programme for flagging potentially fraudulent claims was biased against Black homeowners (allegations, not findings)¹⁴. On 11 September 2023 the court let the disparate-impact theory proceed¹⁴. A February 2024 order limited the first phase of discovery to algorithmic tools, and discovery was still ongoing on 19 February 2026¹⁴. No merits ruling was found. In 2021 Lemonade deleted a tweet saying its AI could "pick up non-verbal cues" in customer videos and called the wording "a bad choice of words"¹⁵.

What was not found. Searches found no regulator enforcement action, audit or withdrawn programme specific to insurance fraud-flagging models. The search was limited, and absence of a finding is not proof that none exists.

Dependence on people and process. The models flag; investigators decide. Savings therefore depend on investigator capacity, data quality and workflow, not on the model alone⁹.

Phase: In early use — why

The hypothesis was In wide use. The evidence supports In early use. Confirmed machine learning use rests on single-insurer reports: Allianz UK's Incognito⁵ and vendor customers⁸⁹. The industry-wide totals cover every detection method, and no page attributes a share to AI²³. Because the scale of genuine AI use cannot be established, the lower plausible phase is chosen. Independent field evidence with a control group, or a regulator figure for AI-flagged claims, would move the phase.

What to watch next

• Huskey v. State Farm: a class certification or merits ruling after discovery, which was ongoing on 19 February 2026¹⁴.
• The NAIC adoption map: the count of 29 jurisdictions as of 8 October 2026¹⁷.
• The ABI's annual fraud release for 2025 data, following releases in September 2024 and November 2025²³.
• Colorado: the Division of Insurance lists claims management as a possible future stakeholder topic¹⁸.
• The final version of the draft UK ICO guidance last updated on 31 March 2026²¹.

Questions to ask before adopting

• Was the vendor's accuracy measured on our own claims, with the share of flagged claims that prove fraudulent, not only overall accuracy?
• Are reported savings net of investigation cost and of delays to legitimate claims?
• Which part of the improvement comes from the model, and which from rules, analysts or shared data?
• Were flag rates tested across customer groups and postcodes, and who reviews the result?
• Who decides after a flag, and can the customer contest the outcome?