TL;DR
Since the first of January a standard exclusion has been sitting inside commercial general liability policies, and it takes artificial intelligence out of cover. More than sixty property and casualty carriers filed a version of it. State regulators approved over eighty percent of the requests. Nobody held a press conference. The retreat is not the interesting part. The interesting part is the market forming next to it, where AI cover exists but the price depends on passing an audit. Insurance is about to do to AI what it once did to electricity.
A clause with four words in it
W.R. Berkley's proposed endorsement bars coverage for any claim involving "any actual or alleged use" of artificial intelligence.
Read it again. Not any harm caused by AI. Any claim that involves its use, actual or alleged. Four words move the burden before the argument even starts.
That clause is not an outlier. The Insurance Services Office, whose templates most American carriers start from, put a generative AI exclusion into commercial general liability with effect from 1 January 2026. CSIS counts more than sixty property and casualty providers filing AI exclusions into their 2026 policies, including Berkshire Hathaway, Chubb, Travelers, AIG, Tokio Marine, W.R. Berkley, Great American and Fairfax. On 23 April 2026 The Information reported that state insurance commissioners had approved over eighty percent of those requests.
No hearing, no announcement. Fifty separate dockets and a form.
What underwriters know and vendors do not say
A refusal to price something is the most honest statement about its reliability you can get for free.
A vendor benchmark tells you what a system does on a good day. An underwriter has to estimate the tail, put capital behind the estimate, then live with being wrong for thirty years. When eight large carriers decide at the same time that they cannot model a risk, that is not timidity. It is a measurement.
The academic side agrees, and that is the part buyers should sit with. Quanyan Zhu at NYU published Insurance of Agentic AI on 3 June 2026 and says it flatly: the actuarial foundations remain under development, and publicly available insured loss data specific to agentic AI remains limited. There is no loss history because the deployments are barely eighteen months old and most of the failures are still being filed under the word pilot.
Nearly 75 percent of companies plan to deploy agentic AI within two years. 21 percent report a mature governance model for agents.
Those two numbers come from the same Deloitte survey of 3.235 leaders across 24 countries. The gap between them is exactly the space an insurer is being asked to underwrite.
The 1894 answer
This has happened before, and the fix did not come from a parliament.
Fire insurers in the 1890s had the same problem with electricity. They could not price a wire. So in 1894 William Henry Merrill opened a laboratory on the third floor of a fire insurance patrol station in Chicago with 350 dollars of equipment, funded by the Chicago Fire Underwriters' Association and the Western Insurance Union. It was the Underwriters' Electrical Bureau, later Underwriters Laboratories. The mark on the back of your phone charger is named after insurance men who did not want to write a policy blind.
The same machinery is being assembled now, fast and without ceremony. Armilla writes cover for hallucinations and model drift through Lloyd's syndicates, up to 25 million dollars. Munich Re's aiSure settles parametrically against accuracy thresholds, latency floors and drift bands that you have to specify in the contract. AIUC prices in tiers against its own AIUC-1 audit, assembled from the NIST AI Risk Management Framework, the EU AI Act and MITRE ATLAS. Pass the audit and pay less. Fail it and pay more, or get nothing.
Certification is arriving through the premium, not through the statute.
Why this matters for your business
The AI Act tells you what is forbidden. Your underwriter tells you what is affordable. The second constraint binds earlier and bites harder, because it shows up at renewal with a number attached.
So the question for the next board meeting is not whether your agents are compliant. It is whether your current policy still covers them, and what an auditor would need to see to cover them again.
Read the endorsements. That is where this year's real AI regulation was written.