Insurers are writing AI liability policies with caps as low as $50 million, while the realistic exposure from a single widely deployed model failure runs into the tens or hundreds of billions, according to underwriters including Munich Re as reported by the Wall Street Journal.
The gap between those two numbers does not disappear. It sits with whoever deployed the system.
Why underwriting this is genuinely hard
Insurance works on uncorrelated risk. Houses do not all burn down on the same night, so premiums from many policyholders cover losses for a few.
Model failures are correlated by construction. If one widely used model has a flaw, every customer using it has the same flaw at the same moment, across every deployment. That is not a fire, it is a flood — and floods are the risk category private insurers have always struggled to price.
There is no actuarial history either. Underwriters price car accidents from a century of claims data. Nobody has a century of frontier model deployments, and the systems change faster than any loss history could accumulate.
What it means if you ship AI features
Read your policy for the cap and the exclusions, because "AI-related" is doing a lot of work in those documents and it is often defined narrowly. Then assume the remainder is yours.
This is also why vendor indemnities have become a negotiating point in enterprise AI contracts — buyers have worked out that the insurance market will not absorb this, and are trying to push it back up the chain to the model providers.
Sources: Wall Street Journal