Banner: Cyber insurance was not built for AI. A Market Moves piece from IQ4hire.
AI governance · Part 4

Cyber insurance wasn’t built for AI. It’s being rewritten right now.

Insurers are narrowing AI coverage before AI losses arrive at scale. What changed in January 2026, what to ask your broker, and why disclosure cuts both ways.

Part 4 of the series AI governance. Each part stands alone.

Insurers are narrowing AI coverage now, before AI losses show up at scale. Do not assume a loss caused by AI is covered. Ask your broker, in writing, which of your policies pick it up and which exclude it.

Parts 1 to 3 covered what you pay for, how hard to watch it, and whether you can govern it. This part asks the financial question: when you watched correctly and it still went wrong, does anyone pay?

Every new category of risk takes insurance through the same two phases. First comes ambiguity. Nobody knows what is covered, and policies are silent because the new thing did not exist when the language was written. Then comes resolution, and the resolution is never "we will widen the policy to be safe." It is narrower language, new exclusions and sharper questions at renewal. Cyber insurance went through this in the 2000s and 2010s. AI is going through it now, faster.

The tell is in the timing

In January 2026, Verisk's ISO program released optional generative AI exclusion endorsements for commercial general liability policies, including forms CG 40 47 and CG 40 48. ISO writes the standard policy language much of the US industry adopts. Because the forms are optional, one client can have an AI exclusion with one carrier and none with another. Trade reporting says carriers including Berkshire Hathaway, Chubb and Travelers have sought state approval for AI exclusions, and that some approved exclusions began attaching to policies as early as January 2026.

As far as I can find, that is not a reaction to a wave of AI losses, because there is no wave at scale yet. It is insurers moving ahead of the loss data, which is the opposite of how insurance normally works. Usually carriers wait for claims history before they tighten language. Here they are tightening first.

The exclusions are not always labeled

Law-firm and broker commentary points out that AI exclusions are not always labeled as such. Some sit in extensions to existing cyber exclusions, and many live in the endorsements section of the policy, not the main text. Reporting also says the most aggressive activity is in professional liability and errors-and-omissions coverage, while cyber policies remain, for now, the more stable source of cover for AI-related risks. Cyber alone is unlikely to give you a complete answer.

In coverage discussions, one question keeps coming up: was AI the vector, the delivery mechanism for an otherwise ordinary loss, or the peril, the actual source of the loss? It sounds technical. It is also convenient for a carrier, because it offers two ways to argue a claim does not fit, depending on which costs less.

Disclosure can become the exit

Carriers increasingly ask policyholders at renewal to describe their AI use: which systems, which data, how much autonomy. That looks like ordinary diligence. It also builds the record a carrier can use later: we asked, you did not fully disclose, so we are not paying. Broker commentary advises documenting your AI use and disclosing it when asked, because undisclosed AI exposure can give a carrier grounds to deny a claim.

The Air Canada case shows the underlying problem. In 2024, a British Columbia tribunal ruled in Moffatt v. Air Canada that the airline was responsible for what its website chatbot told a customer about bereavement fares. The chatbot described a refund route that the airline's policy did not allow, and the airline had argued it was not liable for its chatbot. The tribunal rejected that. The award was small, but the lesson is large: the loss never fit inside language written for network intrusions and data breaches, and nobody had to write an exclusion for it to be uncovered.

Where it lands

Insurers tighten language before the losses arrive, they ask for detailed AI disclosure at renewal, and exclusions often sit in endorsements. Coverage will arrive in some form, the way cyber insurance eventually became a mature product. The open question is how much gets excluded in the meantime, before anyone outside the industry notices the door closing.

Before your next renewal, list the AI systems you run, what each can decide, and ask your broker in writing which policy covers a loss from each one. Part 5 closes the series with the case for governing the decision. If you cannot yet answer what your stack is allowed to decide, what happens when it is wrong, or who answers for it, that is the conversation an AI diagnostic starts.

Sources

← All Insights