You ask your broker whether your liability insurance covers whatever the AI agent talking to your customers does. They answer with a pause and one sentence: “I’d have to look into it.” That pause is the news. Your agent has been in production for months; the wording of your policy hasn’t changed in years.
The thesis in one line: liability insurance for an AI agent isn’t a question of whether cover exists, but of whether anyone has read the policy since the agent started acting. By default it either says nothing about AI or excludes it, and either way you find out after the loss.
Liability insurance and AI agents: what your policy says, and above all what it doesn’t
For years most policies never mentioned artificial intelligence: they neither covered it nor excluded it. The industry calls it “silent AI”, by analogy with the “silent cyber” problem that blew up after the 2017 NotPetya attack; that’s how the broker Gallagher describes it in an analysis from May 2026. Silence looks like an advantage, but it’s a bet: if there’s a loss, an adjuster decides the cover and, if it comes to it, a court, not a clause you’ve read.
That the agent is your responsibility toward the customer isn’t a hypothesis. In 2024 a British Columbia tribunal ruled that Air Canada was liable for what its chatbot told a passenger about a fare: the company can’t plead “it was the bot”. What many companies still haven’t settled is who pays when “it was the bot” gets expensive.
Where does a loss come from with an AI agent?
- A wrong answer with a financial effect. The agent promises a refund, a deadline or a condition that doesn’t exist, and the customer acts on it. That’s the Air Canada case: a pure economic loss, with no injury and no damaged property, and it’s worth checking whether your cover includes exactly that kind of damage.
- An irreversible action in a system. It deletes, pays, sends or edits a record. The damage is yours or belongs to a third party whose data you handle, and which policy responds depends on who suffers it.
- Data that leaves where it shouldn’t. A permissions failure or a badly chosen external tool. The loss comes in through cyber and data protection rather than liability, and can fall between two policies.
What insurers have changed and why it matters to a European small business
The documented shift comes from the US market. According to Gallagher, ISO—the body that publishes the standard forms for general liability policies in the United States—filed clauses in 2025 that exclude damages arising from generative AI output: form CG 40 47 excludes bodily injury and property damage as well as personal and advertising injury; CG 40 48 excludes only the latter. They’re optional: each insurer decides whether to adopt them. The same source says three insurers were seeking approval to exclude AI losses from their professional and general liability policies.
There’s one more number, with its fine print. In its own survey of insurance professionals, Gallagher reports that one in five say their insureds have already had AI-related losses or claims; of those, a little over half were fully covered, 44% partly covered and 3% not covered at all. It’s the broker’s own survey with no published sample size: treat it as a signal, not a market statistic.
That third phase has already begun. According to El Ecosistema Startup in October 2026, some insurers and startups now offer AI-specific policies, and it mentions Munich Re’s aiSure product. There are no reliable prices published, so nothing below depends on them.
The five questions to put to your broker this week
- Does my liability policy mention AI, either in the cover or in an exclusion? Ask for it in writing, with the policy form reference. “It doesn’t mention it” is also an answer, and you want it signed.
- Does it cover the error of an agent acting on its own, not just that of a person using a tool? That’s the difference between assisting and deciding; many wordings were written with the former in mind.
- What about professional indemnity and cyber? The same error can brush against three policies and fall outside all three. Ask them to cross-check.
- What documentation will I be asked for if there’s a loss? Reasonably, they’ll want to know what the agent could do, who supervised it and what was logged.
- Is there AI-specific cover, and what would it require of me? Not to buy it now, but to get the list: that list is your governance roadmap.
The evidence the market asks for is what you should already have
According to the Aon report cited in that same article, insurers want to see documented AI governance and human oversight. It isn’t an exotic requirement: it’s the work that reduces the risk before you insure it. Start with what the agent can touch—what permissions to give an agent—, move on to how much it decides unsupervised—an agent’s levels of autonomy—and put in writing who answers for each change, as explained in governance and control of AI automation.
What weighs most after a loss is being able to reconstruct what happened. That’s why traceability of AI decisions is, besides good practice, a strong argument in front of an adjuster, and human oversight of AI at scale is what proves “a person reviewed it” wasn’t just a line in the sales deck.
When do you not need to change anything?
If your agent only prepares drafts that a person reviews and signs, or works with internal information with no outward effect, the risk of a third-party loss is low and your current policy probably suffices: the first question, in writing, is enough. And buying a specific policy before you have governance makes no sense either: you’d pay for a risk you can’t yet measure. All of this is operational judgment, not legal or insurance advice: your broker confirms actual cover by reading your policy.
The line to take away: don’t ask whether your agent is insured; ask what your policy says about what your agent does. The first question has a comfortable “yes”. The second has a document, and that document is the one that gets read on the day of the loss.