Insurers Can Write $50 Million of AI Coverage. Underwriters Say the Exposure Runs to Tens of Billions.

The fastest-growing legal liability in the world is running into one of the oldest constraints in finance: insurers will not — or cannot — cover it. The Wall Street Journal reported this week that insurers worry claims from AI failures could outpace the limits of existing coverage, and the numbers behind that worry are stark. Munich Re says its policies covering AI model errors carry a cap of $50 million. Underwriters estimate the potential legal liability of the largest AI companies in the tens of billions of dollars — an exposure estimate, not an assessed claim, but the gap between those two figures has to land somewhere.

Testudo, a UK-based specialist AI insurer, told the Journal that large AI firms are willing to buy billions of dollars of cover and have the money to pay for it — but the market simply cannot provide capacity at that scale. AI companies typically carry general liability and cyber policies today, and the industry insiders the paper interviewed fear that growth in AI-related claims could exceed those policy limits, leaving developers and deployers holding the difference on their own balance sheets.

The deeper problem is that most AI risk is not even priced — it is silent. Broker Aon estimates that more than 90 percent of AI exposure under existing policies falls into silent language: the contract neither explicitly covers nor explicitly excludes losses caused by AI. That ambiguity spans loss types that insurance normally keeps in separate boxes — model errors can cause bodily injury, property damage, copyright disputes, data breaches or fabricated outputs, and no traditional policy was drafted with all five in mind. The Insurance Services Office, which writes the standard clauses much of the industry uses, rolled out optional generative-AI exclusions this year and is considering separate ones for agents.

Not everyone wants the exclusions. Marsh, one of the world's largest brokers, wrote to major insurers this month opposing their broad adoption, arguing that blanket AI carve-outs increase uncertainty for policyholders and invite coverage disputes. The broker acknowledged that autonomous systems create genuinely novel risks requiring joint attention, but said that if insurers insist on AI exclusions, it will advise clients to take their business elsewhere. The market is thus splitting along a new axis: insurers that price AI risk, insurers that exclude it, and a lot of policyholders who do not yet know which kind they bought.

Underwriting AI is hard for a structural reason, not just a legal one. Insurance works on uncorrelated risk — houses do not all burn down on the same night. Model failures are correlated by construction: if a widely deployed model has a flaw, every customer has the same flaw at the same moment. That is not a fire; it is a flood, the risk category private insurers have always struggled to price. And there is no actuarial history to lean on, because nobody has a century of frontier-model deployments, and the systems change faster than any loss history could accumulate.

The consequences are already visible in contracts and valuations. Enterprises are negotiating vendor indemnities harder, knowing the insurance market will not absorb the risk, and deployment decisions increasingly weigh what a company's policies actually say about AI. Previous reporting on rogue-agent claims noted that executives themselves could face suits under directors-and-officers policies — meaning the liability chain runs all the way up to founders. If AI losses keep outgrowing insurer capacity, the industry's phenomenal growth rate becomes its own insurance problem: the more the economy runs on a handful of models, the more a single flaw looks less like an accident and more like a catastrophe no one can transfer.