Let’s talk about the one page in your renewal pack that quietly decided how much AI your business is allowed to use.
It doesn’t look like a decision. It looks like admin. A PDF from the broker, forty pages, your name spelled wrong at the top, and somewhere around page twenty-six a schedule of endorsements with form numbers on it. CG 40 47. CG 40 48. CG 35 08. Three lines of reference code that most people scroll straight past on the way to the premium.
Those three forms took effect on the first of January this year. They remove coverage, under a standard commercial general liability policy, for bodily injury, property damage, and personal or advertising injury arising out of generative AI.
Read that again slowly, because it is doing more work than it appears to.
Now picture the meeting. Finance director, ops lead, the broker on a screen, renewal due in nine days. Somebody asks whether the policy still covers the thing they rolled out in March — the AI assistant that drafts customer responses, the model that reads claims documents, the copilot the sales team has been using since nobody can remember. The broker says he’ll check. Two days later the answer comes back in an email with a screenshot: excluded. Not “excluded unless you buy the endorsement.” Excluded.
And the room goes quiet, because nobody in that room can answer the question the underwriter actually asked. Which tools? Which models? Used for what? Who reviews the output? What happens when it’s wrong? There’s no list. There was never a list. Forty SaaS products added an AI feature in the last eighteen months and not one of them sent a letter explaining what changed underneath.
That silence is the whole story.
Here’s what actually happened, and it happened faster than most people noticed. On the twenty-third of April this year, The Information reported that US state insurance commissioners had approved more than eighty percent of the AI exclusion requests filed by Berkshire Hathaway, Chubb, Travelers and AIG. Not a handful. The overwhelming majority, waved through quietly, state by state, in a process that gets about as much press as a zoning variance. By July, more than sixty property and casualty groups had filed to adopt AI exclusions. Berkley went further with form PC 51380 — an absolute AI exclusion that reaches directors and officers liability, errors and omissions, and fiduciary cover.
Now, the easy read is that insurers got scared. That’s not it. Insurers don’t get scared, they get selective. What happened is that generative AI failed the test they’ve used for forty years to decide whether a risk can be insured at all — and it failed on the criterion that matters most. Information asymmetry. The carrier cannot see which models you’re running, what you’re running them on, how the system is governed, or whether the controls you described in the proposal actually exist. According to a 2025 Geneva Association assessment, AI fails three of the nine criteria outright and is under strain on five more.
A carrier that can’t see the risk can’t price it. And a carrier that can’t price it has one move left: take it out of the policy and hand it back to you.
The CSIS report on this in September put it in a line I haven’t been able to get out of my head. Insurance has quietly become the most important de facto regulator of AI deployment in America, and it is currently regulating by withdrawal.
Sit with the shape of that. No consultation paper. No transition period. No regulator to appeal to, no hearing, no reasons given. Just an endorsement, an effective date, and a claim that now belongs to nobody.
Here’s what most operators are missing.
First, this is not aimed at AI companies. It’s aimed at you. The phrase is “arising out of generative AI,” and that phrase doesn’t care whether you build models or just use one to summarise a contract. If your marketing copy was drafted with a model, if your support replies get suggested by one, if a machine helped write the incident report that later becomes evidence, you’re inside the exclusion. The company that thinks of itself as a logistics firm or a law firm or a manufacturer is exactly who this catches, because it never thought to check. The AI vendor has a legal team reading its own contracts. You have a broker who’s behind on his emails.
Second, the cover doesn’t disappear cleanly. It fragments. Cyber narrows in one place, tech E&O narrows in another, D&O gets an absolute exclusion, employment practices liability quietly changes its definitions, and each line points at the next one. Fenwick’s read on this is blunt: it isn’t always a big conspicuous exclusion, it’s revised base forms and narrowed definitions and carve-backs that erode cover without anybody being able to point at the moment it happened. What that produces is a claim with no home. And a claim with no home doesn’t disappear — it goes on the balance sheet, or it goes onto the personal exposure of the director who signed the deployment off.
That’s the part that should make a board sit up.
Third — and this is the piece I’d argue is being almost entirely missed — this is the regulation that actually binds. Parliaments take years. The EU AI Act took years, and half its obligations are still phasing in, and the high-risk rules that went live in August are already being argued about by people who’ve read the annexes and nobody else. An insurance endorsement took a quarter. It applies in dozens of states at once. It has no political constituency to fight, no opposition party to embarrass, no press conference. And unlike a statute, it doesn’t need to prove anything in court to change your behaviour. It just has to be in the renewal.
Which means the most consequential AI policy decision in your company this year was probably made by an underwriter in another country who has never heard of you.
So what do you actually do about it. Four things, in order.
Get the list. Every tool, every model, every place a generated sentence touches a customer, a patient, a payment, or a legal position. If you can’t produce that list in an afternoon, that’s the finding — not a gap in the paperwork, a gap in control. The underwriter is asking the question because he can’t answer it himself, and the honest answer is the one that gets you a conversation instead of a decline.
Read the endorsements, not the summary. Insuring agreements, exclusions, definitions, endorsements, together, in one sitting, with somebody who does this for a living. Ask specifically whether you have an affirmative grant of AI cover, in writing, and what it’s conditioned on. “It should be fine” is not a coverage position.
Build the evidence file before you need it. Model documentation, who approved the deployment, what human reviews the output, what the logs show, what you do when the system fails. This is the boring stuff the AI Act asked for and most companies treated as a compliance tax. It turns out it’s also the underwriting file. The companies that can hand an underwriter a real governance record will get priced. Everyone else gets excluded and told it’s a market-wide thing.
And price the gap. If the risk is uninsured, it’s a cost. Put a number on it, decide whether you’re big enough to carry it, and if you’re not, look at manuscript endorsements, captives, difference-in-conditions layers. That’s not exotic. That’s what every industry did when cyber went through the same door twenty years ago.
Because here’s the pattern underneath all of it, and it’s the reason I think this matters beyond insurance.
The market will only insure what it can verify. That is the oldest rule in underwriting and it has not changed. Which means the thing regulators keep failing to mandate — proof of how the machine was used, by whom, under what oversight, with what record — is about to be demanded by the only party with real leverage and no need to hold a press conference.
Verifiable governance stops being a cost centre and becomes a licence to operate. Not because a politician said so. Because nobody will write the policy without it.
And that flips the whole AI conversation. For three years the argument has been about capability — which model, which benchmark, which agent framework. The binding constraint was never capability. It’s whether a human being can put their name to the deployment and still be standing if it goes wrong. Right now the answer, for a lot of companies, is no, and they don’t know it yet.
The human in the loop was supposed to be the safeguard. At the moment he’s the only uninsured thing in the chain.
That’s the bit worth fixing. Not because it’s fair — it isn’t — but because a system where the person who signs carries all the downside and the vendor carries none is a system that will keep pushing decisions onto people who can’t afford to think about them. And a machine that no human can safely say yes to hasn’t been adopted. It’s been smuggled in.
You’ve got this. But read page twenty-six first.