Two pieces of insurance news landed within a day of each other, and together they say something useful about where AI risk is heading. On September 23 Travelers published its latest Risk Index, in which cyber threats ranked as the top overall business concern for the fifth time in eight years. Of the 1,202 US business insurance decision makers surveyed, 58 percent said they worry a great deal or some about cyber risks, and 55 percent specifically flagged a security breach or cyber event involving AI. Meanwhile in London, underwriters are openly discussing limiting coverage for agentic AI risk, because the market cannot yet price it.
Why underwriters are uneasy
Insurance runs on history. An underwriter prices a risk by looking at how often something has gone wrong before and how much it cost. Agentic AI, meaning AI that takes actions rather than just answering questions, has almost no loss history to look at, and the early signals point to attacks moving much faster than the models were built for. When the time between a vulnerability appearing and being exploited at scale compresses from weeks to hours, the assumptions behind a cyber policy start to wobble. A market that cannot price something tends to do one of two things: charge more, or carve it out.
Two caveats worth keeping
The Travelers survey is of US decision makers, so treat the percentages as directional rather than a Canadian measurement. And nothing here says AI coverage is being withdrawn today. Underwriters are considering limits, which in practice means the questions get sharper first and the wording changes later. That is still worth acting on, because the questions are what you have to answer at renewal.
What this means at your renewal
Cyber insurance applications already function as a de facto security audit, which is something we have written about in what a cyber policy actually covers. Expect the AI section to grow. Plausible questions: do you use AI assistants or agents in the business, what systems can they reach, who approved them, do you have a written policy for employee use, and does a human approve anything that moves money or changes access. If your honest answer to most of that is that nobody has really looked, that is now a pricing input, not a philosophical question.
What to actually do
- Write the policy before the questionnaire asks for it. One page covering approved tools, what may never be pasted into them, and who signs off on new ones is enough to answer honestly.
- Inventory the AI tools already in use, including the ones nobody approved. You cannot describe your exposure if you do not know what is connected.
- Check what the agents can reach. An assistant with access to your mailbox, files and calendar is a set of permissions, and it should be scoped like any other account.
- Keep a human on money and access. Payment changes and permission grants should not be fully automated, whatever the tool promises.
- Ask your broker directly whether actions taken by AI systems you deployed are covered, and get the answer in writing rather than inferring it from the policy summary.
- Do not let AI crowd out the basics. Multi-factor authentication, tested backups and patching are still what most of your premium is priced on.
The takeaway
The useful signal here is not that insurers are worried. It is that AI governance has quietly become a commercial question with a number attached. Businesses that can describe what their AI tools are allowed to do will get better answers at renewal than businesses that cannot, and the work to get there is mostly writing things down and scoping some permissions.