October 12, 2026

Your Business Insurance May Not Cover an AI Mistake

Insurers have spent the past year writing AI exclusions into business policies. As AI agents start acting on their own, most owners haven't checked which ones they signed.

By Jay Cadmus

The Short Version

Last week, at an insurance industry forum in Colorado Springs, executives from Zurich, Willis, Coalition and the broker EPIC said something that sounds reassuring. Blanket AI exclusions, they said, aren't catching on, because AI is now part of how almost every business works.

The filing record tells a less comfortable story. A review of state insurance filings by Trades Coverage counted more than 4,000 adoptions of generative AI exclusion forms across 49 states and Washington, D.C., with about 2,400 already in effect by the end of July. Most sit on the kinds of policies small businesses and contractors buy.

That gap matters more this month than it did last month. Anthropic disclosed that its testing agents sent a fake police tip through a public form, and the State Department says they also submitted incomplete visa applications. A White House task force then told AI companies that incident reporting is "not optional," and administration officials have argued that liability lawsuits will keep AI builders in line better than new rules. Lawsuits need someone to pay. For a small business using AI tools, the answer may already be printed on an endorsement page it has never read.

What the AI Exclusions Actually Say

Most commercial insurance in the United States runs on standard forms written by ISO, the policy-drafting arm of the data company Verisk. In 2025, ISO released a set of optional generative AI endorsements for general liability coverage, with a January 2026 edition date. Insurers decide whether to attach them, state by state and policy by policy.

Three forms do most of the work. The broadest, CG 40 47, removes coverage for bodily injury, property damage, and personal and advertising injury "arising out of" generative AI. CG 40 48 removes only the advertising and personal injury part, the coverage that usually answers claims for defamation or copying someone's slogan. CG 35 08 applies to products and completed work, so it reaches a contractor's finished job or a product a business sold.

The definition is wide. The forms describe generative AI as "a machine-based learning system or model that is trained on data with the ability to create content or responses," including text, images, audio, video and code. That covers the chatbot drafting your client emails, the tool that writes ad copy, and the assistant that helps estimate a job.

"Arising out of" is the phrase that should make an owner sit up. In insurance language it usually means a causal connection, without regard to how large AI's role was. Trades Coverage's head of research, Matt Levin, gave the contractor version: a firm might use AI for one estimate, takeoff or design decision and do the rest of the job by hand. If that piece contributes to a later claim, the broadest form could still apply. No court had ruled on that reading as of July, and one coverage lawyer told Insurance Journal she expects the first fights over AI exclusions to come in professional and management liability lines.

Who Uses Them

Here is where the two pictures collide. At the forum, Zurich U.S. CEO Alex Wells said the ISO forms are "generally only being adopted by smaller carriers who automatically adopt things that ISO does." His company reviewed the forms and doesn't use blanket AI exclusions, preferring to handle AI risk when it prices and accepts each account. Willis's Pat Donnelly said he hadn't seen AI exclusions become common in the policies his firm places. Coalition's Shawn Ram said some brokers threatened to move business if insurers imposed them.

Trade reporting points the other way. Insurance Business reported last month that W.R. Berkley, Chubb, Travelers, Berkshire Hathaway and AIG had each filed by April to adopt the ISO forms or their own versions. Berkley has gone furthest, with what it calls an "absolute" AI exclusion on its directors and officers, errors and omissions, and fiduciary products.

Both accounts can be true at once. A filing gives a carrier permission to use a form. It doesn't prove the form is on any given policy. The Trades Coverage count also tallies one adoption per carrier, form and state, so a single insurer using one form in many states shows up many times. The review came from a company that sells insurance to tradespeople, and no outside group has checked it. Verisk itself told Insurance Journal in July that it didn't know how many insurers had adopted the forms.

My read is that large insurers are keeping the option open while underwriting case by case, and smaller regional carriers are more likely to attach the forms as written. If your business buys from a regional carrier through a local agent, that is the group more likely to carry the standard language. The only way to know which group you're in is to look.

Agents Make the Open Question Expensive

All of this was a slow-moving paperwork story until AI tools started taking actions on their own. On Friday, Reuters reported that cyber insurers are reviewing whether autonomous AI systems even fit a policy's definition of an attacker, after CrowdStrike tied a string of attacks on South Korean banks to what it called likely a single person running AI agents. The same report said insurers are working out who bears liability when an AI-generated action causes a loss.

Insurers also reportedly expect multimillion-dollar claims over rogue AI agents, with some of that exposure possibly reaching the executives who run AI companies. And on Saturday, Bloomberg reported that the administration is leaning on liability law to police AI. Treasury Secretary Scott Bessent told Congress last month that "the best way to guarantee safety is that the creators are liable for what they build and generate." Legal experts told Bloomberg it's unclear whether blame lands on the company that built a model or the business that used it.

That second possibility is the one that reaches Main Street. AI Next Wave covered Anthropic's misfiring test agents on October 11. A business that points an agent at real websites, real customers or real money can end up as the party a plaintiff names. Whether its policy responds depends on wording it chose, or accepted, at renewal.

Some insurers are moving the other way. Coalition has offered an endorsement since 2024 that counts an "AI security event" as a security failure and covers wire transfer fraud carried out with deepfakes. EPIC's Marc Kunney warned about a catch, though. Some policies written to cover AI list the specific AI risks they include, which can leave everything unlisted outside the policy. He called those "cloaked exclusions."

A Renewal Walk-Through

Picture Dana, who runs a 12-person landscape design firm. Her designers use an AI tool to draft planting plans and cost estimates, and her office manager uses a chatbot for client emails. Her general liability policy renews in January, and she's thinking about adding an AI agent that books site visits and sends invoices.

Before anything else, she asks her agent for the full forms list on her current policy and searches it for CG 40 47, CG 40 48 and CG 35 08. If any appear, she asks in writing what they would do to a claim tied to an AI-drafted plan. She asks the same of her cyber policy: does funds transfer fraud cover a deepfake voice call telling her bookkeeper to wire money?

Next she writes down every place AI touches client work. That list takes an afternoon, and it does double duty. Underwriters increasingly look for human review of AI output, according to Aon, and a firm that can show sign-off rules may get better terms.

The booking agent waits until those answers come back. If she goes ahead, she runs a 30-day pilot on site-visit scheduling only, with invoicing kept manual. Her baseline is the time her office manager spends scheduling now. She counts only bookings the agent gets right without edits, and she logs the minutes spent checking its work. Where can it fail? A double booking, a wrong address, a message to the wrong client. Her office manager can switch it off on the spot, and Dana decides whether it stays.

If her policy carries the broad exclusion and her carrier won't remove it, she has a decision to make. She can shop the renewal, narrow how AI touches finished designs, or accept the gap with her eyes open. Trades Coverage's review of filings found no comparable standalone product for contractors that covers the excluded exposure, so shopping may be harder than it sounds.

Opportunity Radar

The problem: small businesses use AI everywhere and can't say which of those uses their insurance covers. Brokers can't answer either without an inventory. The offer is a fixed-fee AI coverage map: a two-hour interview that lists every AI tool and what it touches, matched against the client's forms list, with a one-page summary for the renewal meeting. The likely buyer is the independent agency, which could hand it to clients as a retention tool. Test it cheaply by offering ten free maps to one agency's clients before January renewals and tracking whether any change their coverage. Walk away if agencies won't pay, or if carriers start publishing plain-language AI coverage summaries themselves. Demand is a hypothesis until those ten maps come back.

What You Can Do With This

Small business owners

Ask your agent this week for your policy's full forms list and look for CG 40 47, CG 40 48 and CG 35 08. If any appear, get a written answer on what they would mean for a claim tied to AI-assisted work. Do it before renewal, when you still have room to negotiate.

Contractors and freelancers

The products and completed work form is the one to watch, because it can follow a finished job for years. Keep a simple record of where AI touched an estimate or design, and who checked it. That record helps with an underwriter now and with a claim later.

Insurance agents and brokers

Your clients will ask once the agents they deploy start making mistakes. Building an AI-use question into renewal intake now lets you place coverage on facts. It also protects you from a client saying no one warned them.

Anyone rolling out an AI agent at work

Before an agent sends, books or pays anything on your organization's behalf, find out who in finance or risk has read the insurance language. If no one has, that conversation comes first.

The Bigger Picture

Washington is leaning on lawsuits to keep AI builders honest, and insurers have spent a year deciding which of those lawsuits they'll pay for. Those two decisions meet in the policy files of ordinary businesses. The companies building frontier models have lawyers and risk teams to fight over who is liable. A landscape firm, a dental office or a three-person marketing shop has an endorsement page and a renewal date. As AI agents take on more tasks that touch customers and money, the cost of an AI mistake will settle wherever the policy language leaves it. For most small businesses, the cheapest protection this year is one hour with their policy before January.

References

Business Insurance, "AI exclusions struggle to gain traction," Gavin Souter, October 7, 2026 (updated October 8) Independent Insurance Agents & Brokers of America, "Verisk to Roll Out New General Liability Exclusions for Generative AI Exposures," October 21, 2025 Insurance Journal, "Insurer Interest in AI Exclusions Growing as Risk Becomes Omnipresent," Don Jergler, July 22, 2026 Insurance Business, "The AI clause that may already be on your contractor client's policy," Mark Rosanes, August 25, 2026 PHCP Pros, "New Trades Coverage data shows rapid adoption of AI exclusions in business liability insurance," August 27, 2026 Insurance Business, "ISO's generative AI exclusion is already on thousands of CGL policies," Mark Rosanes, September 16, 2026 MoneyGeek, "Your Business Insurance May Have Dropped AI Coverage in January 2026," Myryah Irby, August 17, 2026 Reuters via Claims Journal, "South Korean Banks Likely Hacked By China-Based Actor With AI Agent, CrowdStrike Says," October 9, 2026 PYMNTS, "AI CEOs Could Be Held Liable For Rogue Model Actions," October 6, 2026 The Next Web, "Trump's AI liability push leaves rogue-agent blame unclear, Bloomberg says," October 10, 2026 Axios, White House Super Intelligence Force statement on Anthropic incidents, October 9, 2026 Coalition, "Coalition Adds New Affirmative AI Endorsement to Cyber Policies," March 26, 2024