AI Insurance: When Emerging Risks Outpace Claims Data

AI has become the leading emerging risk among experts surveyed by AXA. Yet the insurance industry faces a fundamental question: how do you estimate the financial consequences of a technology when reliable historical claims data remains limited?

The difficulty is not necessarily that AI-related losses are absent. Existing insurance records may not consistently identify where AI contributed to an incident, which systems were involved or how their behaviour influenced the outcome.

For insurers, this creates uncertainty around exposure, frequency and severity. For companies deploying consequential AI, it raises another question: what evidence can they provide about their risks before a mature claims history exists?

AXA's 2026 Report Shows How AI Risk Perception Is Changing

In its Future Risks Report 2026, published in October, AXA ranked artificial intelligence and big data as the leading emerging risk among surveyed experts, ahead of climate change.

AI had ranked fourth in 2025.

Conducted with Ipsos, the research surveyed 4,345 experts across 62 countries and regions, alongside 24,000 members of the public in 19 countries. Participants evaluated major risks according to their potential societal impact.

Among experts ranking AI among their leading concerns, 54% regarded it as a rapidly emerging risk. Only 9% believed public authorities were well prepared to address it.

These findings reflect growing concern about AI's potential consequences. They do not measure insured losses or establish that AI-related claims have increased.

That distinction matters because expert risk assessments and insurance claims answer different questions. Expert assessments consider possible future consequences, while claims records primarily document losses that have already occurred, been reported and fallen within relevant insurance arrangements.

A risk can therefore become more significant in forward-looking assessments before a corresponding pattern becomes identifiable in claims data.

What Coalition's Claims Data Actually Tells Us

Coalition's 2026 Cyber Claims Report, published in March, provides a different perspective: losses already affecting insured businesses.

Drawing on 2025 experience across more than 100,000 policyholders, Coalition reported that business email compromise and funds transfer fraud accounted for 58% of observed cyber claims.

The report also highlighted the continuing significance of social engineering and ransomware, demonstrating how established cyber threats shape insurers' loss experience.

These findings are valuable for understanding the incidents that generate claims, their financial consequences and the effectiveness of existing controls.

However, Coalition's published findings do not establish a separate claims trend for losses directly attributable to autonomous AI agents.

That absence should not be interpreted as evidence that AI-related losses are insignificant or uncommon. The report examines cyber claims within Coalition's insured portfolio, using categories that do not necessarily distinguish AI involvement as a separate causal factor.

For example, an incident involving AI-assisted fraud might be recorded according to the resulting financial loss rather than the technology involved.

AXA and Coalition therefore offer complementary but fundamentally different perspectives. One captures expectations about emerging risks. The other documents historical insured losses within established cyber categories.

Neither dataset independently establishes the current frequency or expected severity of autonomous AI liability claims.

The Insurance Industry Has Acknowledged the Evidence Gap

The challenge becomes clearer in research from the Lloyd's Market Association (LMA).

In its Understanding AI Exposures: AI Loss Scenarios Survey Results, the LMA identified a shortage of underwriting and claims data relating to AI exposures.

To examine the problem, it surveyed market participants about potential losses across professional indemnity, cyber, product recall, and accident and health insurance.

The scenarios included erroneous AI-generated professional advice, system downtime caused by AI malfunction and physical harm involving AI-enabled systems.

Respondents regarded professional indemnity as the most concerning scenario. The cyber scenario involving AI-related system downtime received an average viability rating of 3.42 out of five, within the survey's plausible category.

These were market participants' assessments of hypothetical scenarios, not actuarial estimates derived from actual AI claims.

The LMA also identified a practical obstacle: insurers may lack sufficient information about how policyholders use AI within their operations.

Without that visibility, estimating exposure becomes difficult even before calculating the probability of a loss.

The International Association of Insurance Supervisors' 2025 Global Insurance Market Report reinforces this concern.

Among survey responses concerning monitoring and reporting of AI-related liability underwriting risks, 48.4% indicated limited evidence and 29% indicated no evidence.

These percentages describe the reported state of monitoring and information, not the proportion of insurers lacking claims histories.

Nevertheless, they illustrate the difficulty of developing a comprehensive view of AI liability exposure while insurance products, underwriting practices and supervisory approaches continue to evolve.

Why Historical Claims May Not Capture Emerging AI Exposure

Several characteristics make AI risk difficult to evaluate through historical losses alone.

Attribution is complicated. An operational failure may involve an AI model, human decisions, application design, permissions and external systems. Establishing whether AI caused or contributed to the loss requires detailed investigation.

Deployment histories are limited. Some autonomous workflows have not operated at meaningful scale long enough to produce stable loss patterns. An absence of reported claims does not establish an absence of material exposure.

Systems change. Model updates, new tools, modified instructions and expanded permissions can alter an application's behaviour. Historical experience with one configuration may not accurately describe another.

Losses may be correlated. Organisations using the same underlying model, software component or external service could experience related failures, raising questions about accumulation across insured businesses.

Policy wording and reporting practices also influence which incidents appear in insured-loss records. A loss may occur without generating a claim under the particular policy being examined.

None of these limitations makes claims data less important. They explain why emerging AI risks may require additional information before insurers can estimate exposure with greater confidence.

What Can Be Measured Before Claims Data Matures?

Insurers already use scenario analysis, exposure information, engineering assessments and expert judgement when evaluating risks with limited historical experience.

For AI, these approaches can be supplemented with technical evidence from specific deployed applications.

Consider a technology company using an autonomous customer-service agent. It could assess whether manipulated instructions lead to unauthorised information disclosure or whether defined controls prevent the agent from exceeding its permissions.

An infrastructure agent could be evaluated against scenarios involving unintended system changes or inappropriate tool execution.

Such assessments can document the conditions tested, observed behaviour, safeguards and remaining vulnerabilities.

They cannot establish how frequently a loss will occur across an insured population or independently determine an appropriate premium.

Their contribution is more specific: they provide evidence about the characteristics of a particular deployment and how its controls perform under defined challenges.

For companies preparing for broker discussions, customer assurance reviews or insurance renewals, this makes the description of their AI exposure more concrete.

Instead of relying exclusively on general assurances, they can explain which consequential failure scenarios were examined, what the testing revealed and which risks remain.

How AetherLab Helps Institutions Evaluate AI Exposure

The research highlights a measurement problem. Historical claims provide essential information about realised losses, but they may not adequately describe the behaviour or control effectiveness of emerging AI applications.

AetherLab provides independent AI risk assessment and technical evidence to support institutions evaluating, mitigating and underwriting AI risk.

AdversarialScan evaluates defined behavioural vulnerabilities in generative and agentic AI applications. Within an agreed assessment scope, testing can examine responses to adversarial inputs, instruction manipulation and relevant behavioural boundaries.

Evidence Pack documents assessment conditions, observed findings, mitigation measures, retesting and residual exposure.

For an AI-deploying technology company, this creates a record of which consequential failure scenarios were assessed, how the application behaved, which controls were effective and what risks remain, even when historical claims data for that particular application is unavailable.

The evidence does not replace actuarial analysis, claims experience or underwriting judgement. It contributes a different perspective: what has been observed when a specific AI system's controls are challenged.

The Next Challenge Is Better AI Risk Measurement

AXA's findings demonstrate how rapidly AI has moved up the emerging-risk agenda. Coalition's claims experience shows that established cyber threats continue to generate substantial insured losses.

The LMA and IAIS research identifies the challenge connecting these perspectives: the insurance market still lacks comprehensive information about many emerging AI exposures.

The response should not be to dismiss expert concerns because claims records remain incomplete, or to assume that every emerging scenario will generate significant insured losses.

It should be to improve how exposures are identified, evaluated and documented while claims experience develops.

For companies deploying consequential AI, this means understanding not only what their systems are designed to do, but also how they behave when tested against defined risks.

Explore AetherLab's AdversarialScan to understand how independent adversarial testing can help institutions identify behavioural vulnerabilities, evaluate safeguards and build technical evidence that supports more informed AI risk and insurance discussions.