An AI exclusion is an endorsement that removes claims arising out of artificial intelligence from a policy that would otherwise respond. The narrow ones name generative AI. The absolute ones reach any use of AI at all. This page lists the ones we have verified against a form, a filer’s own release, a regulator’s record, or a dated law-firm or trade-press article that quotes the wording, and it names filers and form numbers because the reference is useless without them. Nothing here implies a relationship between Risklytics and any filer.
What ISO filed
ISO, the Verisk unit that writes the advisory forms most commercial general liability policies are built on, published three optional endorsements for the January 2026 edition. CG 40 47 removes bodily injury, property damage and personal and advertising injury arising out of generative artificial intelligence from both coverages of the general liability form. CG 40 48 removes only the personal and advertising injury piece. CG 35 08 applies the same exclusion to the standalone products and completed operations coverage part. All three define generative AI as a machine-based learning system or model that is trained on data with the ability to create content or responses, including but not limited to text, images, audio, video or code (Independent Agent (Big "I") Virtual University, 2025-10-21). The Coverage B form itself is one page long and adds a single exclusion and a single definition (ISO form CG 40 48 01 26, 2026-01).
Advisory means optional. Each insurer files with each state to adopt or delay. By mid-2026 more than sixty property and casualty groups had filed on AI exclusions, forty-one with at least one subsidiary filing to adopt and twenty filing to delay, and state regulators had approved more than eighty percent of those filings (Gen Re, 2026-08-26). The individual filings behind that count are not published in one place, which is why the table below lists only the ones we have checked.
What named carriers filed
Carrier forms came first and go further. Berkley’s absolute AI exclusion on its D&O, E&O and fiduciary liability products removes claims involving any actual or alleged use, deployment or development of AI, and spells out what that includes: AI-generated content, failure to detect AI-created material, inadequate AI policies, products and services that incorporate AI, chatbot representations, statements and disclosures about AI use, and violations of laws regulating AI (Hunton Andrews Kurth, 2025-05-28). Hamilton’s professional liability exclusion turns on generative AI, defined as any system that produces content such as text, imagery, audio or synthetic data in response to user prompts, and names ChatGPT, Bard, Midjourney and DALL-E as examples (Zelle LLP, 2025-10-31).
State filings show the exclusion moving onto carrier paper: National Union (AIG) generative AI endorsements for a hospice and home health program in Idaho and Illinois, Berkley filings for private-company management liability and crime in Connecticut, and Great American’s updated commercial umbrella and excess forms in Washington. AIG said it has no plans to implement its exclusions (beinsure, 2026-02-01). A filed or approved form is not an applied form. Only the endorsement schedule on an issued policy says which one you have.
Absolute versus limited
The tracker sorts every row by scope. A limited exclusion names generative AI, or a narrower subset, and leaves the rest of the policy alone; the ISO forms are limited. An absolute exclusion reaches any use, deployment or development of AI and, on a D&O form, what the company says about its AI to investors. Policyholder counsel describe the same split. Absolute exclusions have appeared on management liability policies (Fenwick, 2026-06-15), one carrier’s professional liability language is described as arguably more targeted because it names generative tools (Pillsbury Policyholder Pulse, 2026-04-13), and cyber forms remain, for now, the most stable source of coverage for AI-related risks (Fenwick, 2026-06-15).
Which lines carry them matters as much as the wording. General liability and products exclusions answer the physical world, where a robot or a device running a model injures someone. Professional and E&O exclusions answer the financial world, where a model’s output costs a customer money. D&O exclusions answer the investor suit. An AI company can be clean on one line and bare on another with the same carrier.
Generated content versus agentic action
Every exclusion published so far is written around content: a system that creates text, images, audio, video or code, or responds to a prompt. That is the generative definition, and it was drafted with defamation, copyright, privacy and hallucinated advice in mind. An agent is different. It takes an action in a customer’s systems, approves a transaction, moves a machine. Whether a generative AI exclusion reaches an agent’s action is a question the current wording leaves to the claim, because an action is not obviously content or a response, and because a robot’s motion planner is not obviously a model trained to create content.
A trade-press headline reports that Verisk is weighing new exclusions for agentic AI risks (The Insurer, 2026-07-10); the article is paywalled and no form has been published, so it sits in the reported, not yet verified list. When an agentic exclusion is published it gets its own row. Until then, a company whose product takes actions should read the generative definition on its policy and ask the carrier, in writing, whether it applies to the product.
How to negotiate a buy-back or a limited version
Start with the wording. Ask for the endorsement schedule with the quote and read every form whose title contains artificial intelligence or generative. If an absolute exclusion is attached, ask for it to be replaced with a generative-only version, which the ISO forms show is a form the market already has. If a generative exclusion is attached, ask for a carve-back for your own product: language that says the exclusion does not apply to the named product or to claims arising from the technology services the policy is written to insure.
Where the carrier will not move, the market has started to sell the exposure back. An affirmative AI endorsement on a cyber policy adds an AI security event and AI-executed funds transfer fraud to the grant (Coalition, 2024-03-26). An AI liability product for small and medium businesses, distributed through partner carriers, adds bodily injury, property damage and personal and advertising injury from AI use (Munich Re / HSB, 2026-03-18). A cyber product with affirmative AI coverage is reported at limits up to $10 million (Insurance Journal, 2026-04-21). Policyholder counsel also point out that many companies hold multiyear occurrence-based general liability programs that predate these provisions, that courts have found coverage where a complaint mixes excluded and covered allegations, and that an exclusion cannot be read so broadly that it swallows the coverage the policy promised (Pillsbury Policyholder Pulse, 2026-04-13).
We do this on every placement. The submission goes to insurers who cover AI and autonomy explicitly, and a licensed broker reads every endorsement before anything binds.