Insuring AI
Luis Botin, Co-founder9 min read

The question of who bears AI risk is now front and center in the conversation about scaling AI.
What will it take for AI coverage to exist at scale?
Appetite for writing AI risk is nascent. The majority of legal frameworks and insurance products were developed in the context of humans and employees, not autonomous systems. AI risk is new, dynamic, opaque, and highly concentrated, making it difficult to understand and control, harder still to insure. As a result, major carriers lack a validated underwriting model to cover AI risk at scale, excluding for the moment most AI-related losses from their policies.¹
We see the AI insurance market requiring a cross-industry effort towards rigorous testing and clear liability along the chain. As with past technologies, we believe this effort will align incentives towards safe, rapid, widespread adoption of AI.
Barriers to an AI insurance market and what each will require
1. Lack of historical data
Pricing an insurance policy typically requires decades of claims data (you need history to know the probability and severity of claims). In the case of AI, being a new risk with virtually no claims history, standard underwriting breaks down.
Without data, new insurance categories must find other ways to estimate risk.
Cyber risk found itself in a similar position in the 90s (following the commercialization of the internet) when companies sought coverage against hacking and data breaches (a new, little-understood risk at the time). With no claims history, the first underwriters relied on in-depth audits and testing. As the cyber insurance market grew and competition increased, this put pressure towards thinner risk assessments, with many insurers writing policies based on very simple questionnaires. These policies were eventually found to be drastically underpriced, forcing cyber players to raise rates, cut limits, and many leaving the market.² Today, the cyber insurers seen as leaders are pairing coverage with technical assessment of risk (e.g., Coalition, At-Bay), offering bundled security monitoring before and during the policy.
The AI insurance market will similarly rely on audits and testing. However, it starts from a much stronger position. Cyber insurers could assess a company's defenses, but it was (still is) very difficult to accurately model and simulate a risk that is dependent on external attackers, extrinsic to the insured system itself. While some AI risk similarly comes from external attackers (risks such as jailbreak or prompt injection), the majority comes from the system's own behavior, such as a hallucination or an agent taking the wrong action.
The architecture of an AI system (underlying model, data sources, autonomy and guardrails) and its exposure allow for a first estimate of risk. Further, adversarial testing and simulations can reveal how a system fails under different scenarios, while continuous monitoring of the system (post-deployment) provides data on how frequently failures occur. Methods such as red teaming and agent benchmarking produce metrics (e.g., attack success rates, hallucination rates) that serve as proxies for claims frequency in the absence of historical claims data. These act as quasi actuarial data and can serve as a direct pricing input. Additional modeling (e.g., loss distributions, tail risk, correlations between events, uncertainty) completes the risk model.
We believe the market will reward insurers that combine technical, system-based underwriting with a low-friction process for customers. We also believe these will produce the high-quality, time-series, granular loss data needed to price AI insurance in the future. Insurers pricing policies based on thin questionnaires or the little AI incident data available are pricing blind.
2. AI risk is dynamic and constantly evolving
Usually, insurance is underwritten once and renewed every year, which has worked well for traditional assets that change slowly (a building or a vehicle). In the case of AI systems, these are constantly evolving, making them very difficult to assess and price. Within a policy period, the underlying model can update or swap (Opus 5 to 5.5 to Fable), prompts are rewritten, the system is given new tools/permissions, and the data it can access changes. What this means is that the system an insurer assessed and underwrote at the start of a policy may be entirely different a few months later.
One-off assessments will be insufficient. Insurers will need continuous, controls-based underwriting for the duration of the policy. Other insurance lines have already taken a step in this direction; auto coverage relies increasingly on telematics, pricing based on how a car is actually driven (accessible now with Tesla, Rivian, and other connected cars that report driving data directly) instead of on a one-time annual check. AI systems allow more than this. Insurers can verify that the system deployed matches the one they tested (preventing fraud and allowing for more accurate underwriting), monitor its behavior continuously, and retest in the case of important changes (new model version, new tool, permissions). We believe AI coverage will need to operate continuously.
3. Correlated and systemic risk
Insurance models work when losses are (largely) independent of each other (e.g., car accidents or house fires). What this means is that whenever one customer has a loss, it doesn't make others more likely to have one, making the total amount of claims predictable, which in turn allows insurers to set premiums that cover them.
In the case of AI systems, the underlying components are common to almost all companies, which build on the same few foundation models, tech stacks, and cloud providers. Mythos and Fable suspensions⁴ or the Salesloft Drift Breach⁵ hit many companies at the same time, much like the CrowdStrike outage in 2024, where a bad software update paralyzed airlines, banks, and hospitals on the same day. It is difficult to price (and absorb) losses that arrive all at once.
Insurance has dealt with correlated risk in the past. Hurricanes and earthquakes (affecting entire areas at once), for example, are covered through catastrophe modeling, caps on exposure to any one region, reinsurance, and catastrophe bonds, along with government backstops for extreme events (e.g. terrorism). In the case of AI, insurers will need to know which models and providers each company it covers depends on, test their portfolio against a shared failure (e.g. a major model outage or bad update), cap exposure to any single provider, and price the correlation into the policy.¹ Companies can reduce their own dependence by having a second model provider available. In the early days, some systemic risks will likely need to be excluded altogether, so that the remaining risks can be insured viably while insurers learn more about the space. As the market matures, these can be brought back into coverage, with the largest events shared more widely (reinsurance pools, catastrophe bonds, and eventually a public backstop).⁶
4. Opacity and attribution (liability)
Equally important in insurance, following a claim, the insurer needs to know what happened and who was responsible. This decides whether it's the policy that pays, how much, who the insurer can recover from, and so on.
AI makes this very difficult due to model opacity and increasing interpretability issues in frontier labs (this was among Dario Amodei's key concerns motivating We Must Pace the Frontier). It is difficult to explain why a model produced a given output. Models orchestrated into complex agentic systems further cloud attribution, as a failure can trace back to a million different things - the model, training data, prompts, a tool, or the person using the output, each of which belongs to a different company (developer, vendor, deployer).
Attribution is the reason for flight recorders in aviation, for example, allowing investigators to reconstruct incidents and ultimately deciding litigation and insurance outcomes. In the case of self-driving cars (in the UK, at least), the motor insurer pays the victim prior to attribution and recovers from the manufacturer after the fact.⁷
We see a similar approach with AI. AI systems can log each step (model version, prompt, tool calls, human approval) so that a traceable chain of evidence can be reconstructed (even when the model's reasoning cannot be explained). Another discussed solution is parametric insurance, which pays on measurable triggers (such as an error rate exceeding an agreed threshold) and avoids disputes over cause.³ Clear contracts and liability along the chain will be important in order to state in advance who bears which failures.
How we see this new market forming
AI is spreading into the economy, with the latest string of incidents exposing the growing lag of safety infrastructure behind AI risk. Below is our view of the AI insurance market today, and how we see it evolving from here to 2, 5, and 10 years from now.
Today (November 2026)
Liability is shifting from employees to fully autonomous agents. Today, the exposure created by the underlying services provided by third parties clearly outweighs the value of these contracts, creating a big liability on all the end enterprises deploying these agents. Meanwhile, AI losses are either silent or excluded from standard policies (e.g., commercial general liability, professional indemnity (E&O), and cyber insurance).
The AI insurance market is small. A handful of specialists write AI risk with limits of up to $50 million. In its IPO filing, Anthropic told investors they may not be able to buy the coverage they need.⁴
As we discussed in our previous piece, there is currently no skin in the game on AI risk. Nobody bears the cost of an inadequate evaluation of AI. Testing is currently a procedural, regulatory, box-ticking exercise.
Near term (1 to 2 years)
AI is going to be embedded everywhere. Every traditional line of insurance will need to account for this, as a company's risk profile will completely change as operations start to be run by an agentic workforce. Following the latest AI exclusions, standard policy renewals will expose companies to AI losses. This will create demand for dedicated cover.
Insurers able to test and monitor systems accurately will collect the first real loss data.
As with medicine, aviation, and nuclear energy, the requirements for underwriting will become the first safety standards, insurance becoming a market-based incentive for rigorous evaluation of AI systems.
Medium term (3 to 5 years)
There will be two AI insurance markets. The first is AI endorsements across the whole P&C market, which is worth $2.4 trillion in premiums.⁸ The second is standalone AI products for risks that don't fit existing lines, forecasted to reach about $4.8 billion in premiums by 2032.⁹
Common wording, metrics, and incident definitions emerge (making risks comparable), which attracts reinsurers and the first capital markets deals.
Liability becomes clear and is attributed to wherever the risk started in an incident. This allows insurers to recover claims from the responsible party, whether model, agent, or deployment. Anthropic in its IPO filing: "Any determination that we are liable could result in significant liability and require us to change how we design and offer our models, platform, and applications."⁴ Safety becomes a competitive advantage, as loss data gives indication of which controls work. The market (incentive alignment) rewards them.
Long term (5 to 10 years)
AI becomes a major line of insurance. Cyber took 25 years to reach roughly $15 billion in premiums; AI touches more of the economy and is moving faster.
Coverage is embedded in all AI deployments. High-risk uses require it (as with many high-risk technologies).
The largest risks are shared across governments and the markets (as mentioned earlier, public-private pools cover events that are too big for one insurer).
The price of risk is what self-paces and self-regulates the frontier.
Notes
1. Sasha Romanosky and Celine Robinson, The Insurability of Artificial Intelligence, RAND Corporation, 2026.
2. Daniel W. Woods and Josephine Wolff, "A History of Cyber Risk Transfer," Journal of Cybersecurity 11(1), 2025.
3. Lukasz Szpruch et al., "Insuring AI: Incentivising Safe and Secure Deployment," Harvard Data Science Review 8(3), 2026.
4. Anthropic, PBC, Form S-1 Registration Statement, filed with the SEC on October 26, 2026, "Risk Factors."
5. Google Threat Intelligence Group, "Widespread Data Theft Targets Salesforce Instances via Salesloft Drift," August 2025.
6. Alex Leung et al., "The Insurability Frontier of AI Risk," AIFT, June 2026.
7. Automated and Electric Vehicles Act 2018 (UK), c. 18, sections 2 and 5.
8. Swiss Re Institute, "Growing stronger: how the Property & Casualty market adapts to a riskier world," sigma 3/2025, September 2025.
9. Deloitte Insights, “AI insurance could be a $4.8B market by 2032,” video, August 4, 2025.
10. Munich Re, Cyber Insurance: Risks and Trends 2025, April 2025, as reported by Insurance Journal.