Last Friday, almost nobody noticed a statement from Verisk — the analytics firm whose Insurance Services Office (ISO) writes the standard policy language used across much of the US commercial insurance market. As first reported by The Insurer on 10 July, Verisk said it is evaluating additional options to address AI-related exposures — including agentic AI, the autonomous systems that don't just generate content but make decisions and take actions on their own.

That sentence matters because of what ISO has already done. Since 1 January this year, insurers have been able to attach three new optional ISO endorsements to standard US commercial general liability policies — CG 40 47, CG 40 48 and CG 35 08 — the broadest of which excludes coverage for "bodily injury, property damage or personal and advertising injury arising out of generative artificial intelligence." In plain terms: if your company's use of generative AI hurts someone, a policy carrying these endorsements will not respond.

Individual carriers are moving too. W. R. Berkley last year announced what Bloomberg Law describes as an "absolute" AI exclusion, including for directors-and-officers claims. Policyholder lawyers are alarmed at where the logic leads: "Given that everything including our phones with Google now uses AI, I don't know how anything would be left if there is a broad AI exclusion," Reed Smith's Courtney Horrigan, who represents policyholders, told Bloomberg Law this month. Another coverage lawyer, Lathrop GPM's Alana McMullin, warned that "narrow exclusions will still create hidden coverage gaps."

The gap was already there

The uncomfortable truth is that even before the exclusions, conventional policies mostly didn't cover AI harms. The clearest map of this comes from "Smart Systems, Blind Spots", a report the reinsurance broker Gallagher Re published back in March (in association with MIT and Testudo — worth knowing: Testudo is itself an AI-liability insurer, so the litigation data comes from a firm with a commercial interest in the risk looking large).

The report works through the standard corporate insurance stack and finds each layer pointing away from AI-output liability. Cyber policies trigger on network-security failures — a chatbot confidently inventing a refund policy isn't one, so harms like defamation, IP infringement and hallucinations "are generally not covered by cyber policies." Technology errors-and-omissions cover is "a purpose-built product for suppliers of technology" — the vendors, not the far larger population of businesses deploying AI. Product liability responds mainly where AI sits inside a physical product. And the vendors themselves? Their standard terms "typically cap their liability at 12 months of fees" with no performance warranties — leaving deployers holding the risk.

The canonical example is still Moffatt v. Air Canada (2024 BCCRT 149): a British Columbia tribunal held the airline liable for negligent misrepresentation after its website chatbot wrongly told a passenger he could apply for a bereavement fare retroactively — rejecting Air Canada's argument that the chatbot was somehow a separate entity responsible for its own words. Or Zillow, whose home-valuation algorithm went so wrong the company reduced estimated home values by more than $500 million — no breach, no theft, no biased employee, and, as the Gallagher Re report puts it, no traditional policy that would respond.

The numbers — and the honest wrinkle

The report counts more than 700 US lawsuits involving generative AI filed between 2020 and 2025, with filings up 978.1% from 2021 to 2025 (the data comes from Testudo's litigation tracker — see the caveat above). But here is the wrinkle the scary headline hides: the litigation is not mostly about chatbots gone rogue. Patent claims make up 11.9% of cases, copyright 11.2%, and privacy-related personal-injury claims 10.2% — while Testudo's own market overview puts hallucination suits at just 4.9% of the total. The AI lawsuit wave, so far, is dominated by fights over how models were built and trained — the territory of the news organisations' evidence fight with OpenAI and the music industry's labelling push — not customer-facing AI mistakes.

That cuts both ways for buyers of insurance: the everyday-hallucination risk is (so far) smaller than vendors of new AI policies imply, but it is also precisely the category that standard policies were never built to cover — and that the new exclusions now carve out explicitly.

Selling the net back

Where general insurers retreat, specialists advance — and much of the new market runs through London. Munich Re's aiSure offers performance warranties on AI systems. Armilla, backed by Chaucer and Axis Capital, launched in 2025; Testudo, a Lloyd's coverholder, launched in January 2026 targeting enterprises deploying generative AI. AXA XL has bolted a generative-AI endorsement onto its cyber product across the US, Canada, UK, Europe and Asia, and Hiscox UK has added affirmative AI coverage to its technology professional-indemnity policies. As Beazley's Paul Bantick told Bloomberg Law: "You're going to see a lot of innovation around ensuring policies evolve to cover new AI exposures."

For UK and European readers there is a regulatory clock running under all of this: the EU's new Product Liability Directive, which explicitly reaches software and AI, must be implemented by member states by 9 December 2026 — widening the liability funnel just as the insurance that used to sit beneath it is being redrawn.

The direction of travel is clear enough. Generative AI was excluded from the standard forms in January. The standard-forms body is now studying agentic AI — the systems every vendor is racing to sell in 2026. If your business is deploying agents, the time to read your liability policy's endorsements was probably yesterday.