What problem does it solve?
Fraud hides among honest claims. Most suspicious claims look normal when viewed alone: a slightly inflated invoice, a staged accident with credible witnesses, a repair shop or clinic that appears in too many claims, a photo reused from another insurer. Traditional detection relies on business rules and on handlers noticing something odd, which produces many false alerts and misses the organised schemes that span claims and insurers.
The pressure grows as insurers speed claims up. Faster payment and straight through processing are what customers want, and exactly what fraudsters exploit. Special investigations units are small, so the quality of each referral matters more than the number of alerts.
- The Coalition Against Insurance Fraud states that insurance fraud steals at least USD 308.6 billion every year from American consumers and that fraud occurs in about 10% of property and casualty insurance losses.Insurance Fraud Statistics: $308.6B Stolen Every Year (2026)
How does it work?
- Score at first notice of loss. Each new claim is scored in real time so honest claims can go straight to processing and suspicious ones are held before payment.
- Combine many signals. Models use claim and policy history, the text of notes and documents, images, and external data such as industry databases and public records.
- Look across claims. Network analysis links people, vehicles, addresses, repairers and providers across claims and, through industry initiatives, across insurers.
- Explain every alert. Each alert states the scenario and the facts behind it, so a handler or investigator can decide quickly whether to refer, investigate or clear it.
- Keep watching. The score is recalculated as new information arrives, and investigation outcomes feed back into the models.
- Audience
- Back office
- Autonomy
- Assist
- Adoption
- Mainstream
- Channels
- API and system to system, Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Fraud losses prevented | Not pooled | at least EUR 12 million | 1 | 1 vendor |
| Interactions handled | Not pooled | at least 1 million | 1 | 1 vendor |
Value drivers: Risk and loss reduction, Lower cost to serve, Speed and cycle time, Employee productivity.
Indicative value
A property and casualty insurer paying USD 1 billion in claims a year
USD 1 million to USD 6 million
Additional fraudulent payments avoided per year
How this is calculated
Formula: claimsPaid * fraudShare * additionalStopped. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Claims paid per year claimsPaid, USD per year | 1,000,000,000 | 1,000,000,000 | The reference insurer. |
| Share of claims losses affected by fraud fraudShare, fraction of claims paid | 0.05 | 0.1 | The high value follows the Coalition Against Insurance Fraud statement (a US figure) that fraud occurs in about 10% of property and casualty losses; the low value is an editorial assumption. Source |
| Share of that fraud additionally stopped thanks to AI detection additionalStopped, fraction of fraudulent losses | 0.02 | 0.06 | Editorial assumption; replace with results from a controlled pilot on your own book. |
What it leaves out: Avoided fraudulent payments only. It leaves out investigator time saved by better referrals, the faster payment of honest claims, the deterrent effect, the cost of investigations and the cost of the platform. The share of fraud stopped varies widely by line and market.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Lemonade
United States · Insurance · 2025
Lemonade's claims bot AI Jim takes the first notice of loss in a chat with the customer, pays or declines simple claims within seconds and assigns the claims it may not settle, or has concerns about, to human claims experts based on their specialty, workload and schedule. A separate system, Forensic Graph, uses machine learning to predict, detect and block fraud across the customer engagement. The annual report states that AI Jim took the first notice of loss without human intervention 96% of the time and that roughly 55% of claims were automated from start to finish, both as of December 31, 2025.
- Containment rate: 96%, First notice of loss taken without human intervention, as of December 31, 2025
"AI Jim is our claims bot, and, as of December 31, 2025, 96% of the time, it is AI Jim that will take the first notice of loss from a Lemonade customer without human intervention"
Claimed by: organization - Automation rate: about 55%, Share of claims automated end to end, as of December 31, 2025
"As of December 31, 2025, roughly 55% of our claims were automated, resulting in instant or near-instant processing from start to finish."
Claimed by: organization
Tokio Marine & Nichido Fire Insurance
Japan · Insurance · 2025
Tokio Marine & Nichido Fire Insurance uses Shift Technology's claims intake and claims fraud detection solutions, extended with generative AI that extracts data from structured and unstructured sources such as images and documents. The system highlights the points handlers should check for consistency across estimates, damage photos and claim statements, which makes reviews more efficient and more standardized, including during the surge of claims after large disasters, and it helps detect suspicious claims, which tend to rise after such events. No figures were published.
No outcome disclosed.
AXA Switzerland
Switzerland · Insurance · 2023
AXA Switzerland checks motor and property claims for fraud in real time at first notice of loss with Shift Claims Fraud Detection, using more than 100 fraud scenarios tuned to the Swiss market and its portfolio, and combining its own policy and claims data with external sources such as government records. Honest claims go straight to processing, suspicious ones go to an expert with the full context of the alert, and the models run again whenever new data is recorded on a claim. Shift reports that AXA has analysed more than 1 million claims and stopped over EUR 12 million in fraud.
- Interactions handled: at least 1 million, Claims analysed since deployment
"AXA has now analyzed more than 1 million claims with Shift, and stopped over €12M in fraud, freeing its teams to focus on customer satisfaction and achieve the goal of increasing its presence as #1 in the Swiss market."
Claimed by: vendor - Fraud losses prevented: at least EUR 12 million, Fraud stopped since deployment, cumulative
"AXA has now analyzed more than 1 million claims with Shift, and stopped over €12M in fraud, freeing its teams to focus on customer satisfaction and achieve the goal of increasing its presence as #1 in the Swiss market."
Claimed by: vendor
Assurant
United States · Insurance · 2022
Assurant has used Shift Claims Fraud Detection since 2018 to flag suspicious claims to its special investigations unit, after a trial in which the system identified dozens of fraud cases. Alerts carry the context investigators need, the investigation software was adapted to the unit's workflow, and the models were refined for schemes involving specialty vehicles, extreme weather and multiple claims. Shift reports that the unit's case acceptance rate rose over four years, leading to more fraud mitigated, but gives no figures.
No outcome disclosed.
General Insurance Association of Singapore
Singapore · Insurance · 2017
Member insurers of the General Insurance Association of Singapore analyse travel and motor claims together with Shift Technology's AI, which finds connections between people, providers and claims that look genuine when each insurer sees them alone. The data analytics initiative started in 2017 with 25 insurers; fraud alerts are issued to members and prompt joint investigations. The public page gives no outcome figures.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Claims, policy and payment history with investigation outcomes as labels
- Claim notes, documents and images linked to each claim
- External data such as industry fraud databases, public records and sanctions lists where lawful
- Scenario definitions agreed with the special investigations unit
Systems to integrate
- Claims management system for real time scoring and holds on payment
- Special investigations unit case management
- Industry fraud data sharing schemes and databases
- Document and image analysis services
- Subrogation and triage models in the same claims flow
Complexity: High
Scoring itself is well established, often through specialised vendors. The effort is in data: joining claims, policy, payment, document and external data, labelling past investigation outcomes, and fitting alerts into the handler and investigator workflow without flooding it.
- 1
Agree what a good referral is
With the investigators, define the scenarios that matter per line of business and what an accepted referral looks like; that becomes the target and the main quality measure.
- 2
Back test on closed claims
Score several years of closed claims and compare alerts with known fraud and with the current rules, at the same number of alerts investigators can handle.
- 3
Score at first notice of loss
Move scoring to the start of the claim so honest claims are not slowed down and suspicious ones are held before payment.
- 4
Put reasons in front of people
Show each alert with its scenario and facts inside the handler's screen and track the decision taken, so feedback is captured for every alert.
- 5
Join industry data sharing
Organised fraud crosses insurers. Industry schemes that share claims data under clear legal bases find rings that no insurer sees alone.
Guardrails
- The AI raises alerts; people decide on refusal, investigation and any report to authorities
- Every alert carries its scenario and supporting facts, and alerts without reasons are not actioned
- Protected characteristics and close proxies excluded from features, with fairness testing across customer groups
- Honest customers are not delayed beyond a set time by a pending alert without a human decision
- Data sharing with other insurers only under documented legal bases and agreements
KPIs to instrument
- Share of alerts accepted for investigation and share confirmed as fraud
- Fraud stopped per period, compared with the pre AI baseline on the same lines
- Alerts per investigator and time to decision per alert
- Payment delay caused to claims that turned out to be honest
- Complaints and appeals linked to fraud holds
Human in the loop
Handlers and investigators review every alert and decide on the next step; no claim is refused on a score alone. Investigation outcomes are recorded and feed the models, and the special investigations unit approves changes to scenarios and thresholds.
Common failure modes
- Alert floods
- Too many low quality alerts teach handlers to ignore them. Tune to investigator capacity and measure acceptance, not volume.
- Bias against groups of customers
- Features such as postcode that stand in for ethnicity or age treat honest customers as suspects. Test outcomes across groups and remove proxies.
- Honest customers punished for speed
- Suspicious claims are held but nobody looks at them, so honest customers wait. Set a service level for every held claim.
- Models that fall behind fraudsters
- Schemes change quickly, for example with AI generated documents and images. Retrain on recent outcomes and add image and document integrity checks.
What are the risks and rules?
EU AI Act
Depends on design
Claims fraud detection by an insurer is not listed in Annex III, and point 5(b) explicitly excludes AI systems used to detect financial fraud from the credit scoring category. Point 5(c) covers only risk assessment and pricing in life and health insurance, so a fraud model becomes high risk when it also feeds those decisions, or when it is used by or on behalf of a public authority to grant, reduce, revoke or reclaim public assistance benefits (point 5(a)). Profiling and automated decisions remain subject to GDPR, including Article 22 where a claim is refused on a decision based solely on automated processing.
Guidance
- Opinion on Artificial Intelligence governance and risk management (European Insurance and Occupational Pensions Authority, Europe). Calls for fairness, data governance, explainability and human oversight of AI systems used by insurers, proportionate to their impact on customers.
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Shows which insurance and public benefit uses are high risk and that fraud detection is carved out of the credit scoring category.
Controls to put in place
- Documented scenarios and thresholds, approved by the special investigations unit
- Fairness testing of alert rates and outcomes across customer groups
- Audit log of every alert, its reasons and the human decision taken
- Service level for claims on hold because of a fraud alert
- Legal basis and data sharing agreements for external and industry data
Frequently asked questions
- How much fraud can AI detection stop?
- Among the deployments on this page, the only quantified result is a cumulative total, not a rate: Shift Technology reports that AXA Switzerland has analysed more than 1 million claims with its real time detection and stopped over EUR 12 million in fraud. Lemonade says only that its fraud system has helped it avoid millions of dollars of potential losses. Results depend on the line, the market and how well alerts fit the investigators' workflow.
- Why score at first notice of loss rather than later in the claim?
- Early scoring lets honest claims move straight to processing while suspicious ones are held before money leaves the insurer. AXA Switzerland chose detection at first notice of loss for exactly that reason, and reruns the models whenever new claim data arrives.
- Can insurers detect fraud rings that span several companies?
- Yes, through industry data sharing. Member insurers of the General Insurance Association of Singapore analyse travel and motor claims together to find links between people, providers and claims that look genuine to each insurer alone.
How to cite this page
Blits.ai AI Use Case Library, "AI for insurance claims fraud detection", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/claims-fraud-detection. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published