This week the insurance market answered the question at the center of this series — who pays when AI causes harm — and it gave two opposite answers at once. Dark Reading reported October 22 that a group of specialty carriers is now selling purpose-built AI cover, while several of the largest carriers in the market are writing AI out of their general liability policies.
On one side: Armilla, Coalition, Corgi Insurance, Cowbell, Munich Re, and Beazley are selling dedicated AI policies, affirmative endorsements, or AI language inside technology errors-and-omissions cover. Beazley's AI and emerging risk strategy lead, Craig Linton, describes dedicated attack-defense policies plus what he calls "AI voluntary shutdown coverage" — for a system that must be taken offline and purged of unwanted data an AI agent brought in, with no conventional malware failure at all. Beazley's application runs dozens of questions across roughly 15 security controls, including operational technology security and PCI security.
On the other side: Chubb, Travelers, Berkshire Hathaway, AIG, and W. R. Berkley are among the carriers adopting Verisk/ISO exclusion endorsements that remove bodily injury, property damage, and personal and advertising injury arising from generative AI from commercial general liability policies. Global Finance, cited by Jolt Journal, counts more than 60 property-and-casualty carrier groups that had filed some form of AI exclusion as of October 8. The ISO endorsement at the center of it, CG 40 47 01 26, took effect in January 2026. Standalone AI liability limits are reported at $2 million to $50 million — real money, but small against the casualty towers beside them.
What it means for carriers
First, the split itself is the signal. A split this clean usually means the risk is real, the loss data is thin, and nobody yet trusts their own pricing. The exclusion camp is not saying AI risk is uninsurable — only that it doesn't belong in a form never priced for it. The affirmative camp is saying: the buyers are already here, and the carrier that learns to underwrite this first owns the category.
Second, the wording is now the product. Cowbell co-founder and CEO Jack Kudale put it plainly: policy wording decides whether a claim lands. Does the policy address non-malicious system failure — the agent that did something wrong without anyone attacking it? Does it cover unauthorized disclosure without external intrusion? Does that loss sit in cyber, in technology E&O, or nowhere at all? And Kudale's top concern is not hallucination or bias. It is supply-chain exposure, extending to secondary and tertiary vendors. That is the question the whole AI stack keeps dodging: your customer's agent runs on someone else's model, cloud, and data. When it fails, whose policy answers?
Third, this is Nº 004 and Nº 006 coming back as a market event. Berkley priced the exposure; Greenberg called for the industry to put its arms around it. This split is what that looks like before standards arrive — the affirmative carriers drafting the standard by underwriting it, the excluders buying time for the real one.
From the operator's chair
One: audit your own silent AI exposure before your next renewal cycle does it for you. Policies commonly renew at year-end, so your commercial insureds are about to discover their AI exclusions at the same time you are. Find every place AI risk already sits in your book unpriced — casualty forms written before generative AI, cyber wordings that assume a human attacker, tech E&O that never contemplated an agent acting alone.
Two: decide, explicitly, whether you are an exclusion carrier or an affirmative carrier — line by line, not firm-wide. You can exclude generative AI from general liability while writing affirmative AI cover in specialty lines, the way the specialists do. What you cannot do is let each underwriting team improvise its own answer and call the aggregate a position. One named owner for AI exposure across the firm, able to say, for each line, whether AI risk is in, out, or priced — and why.
Three: steal the specialists' application. Beazley's control list is the closest thing the market has to an AI underwriting standard today: guardrails, human oversight, supply-chain mapping, and evidence of authorized execution. That is my standing doctrine wearing an underwriter's clothes: agents pre-qualify, a licensed person confirms, the ledger keeps the tape. A carrier that can show that discipline in its own operations can underwrite it in others. A carrier that cannot will keep excluding — and keep watching the specialists take the premium.
The market just split on AI risk. Both halves are rational. But only one half is learning. Pick your side on purpose, and make sure the side you picked is the one doing the learning.
— Krishna