Insurers are bracing for multimillion-dollar claims from "rogue AI" agents, the Financial Times reported this week. According to Fortune's summary of that story, an Aon analysis examined more than 300 AI-related cases and found potential exposure across cyber, crime, intellectual property, technology errors and omissions, and other policies.

The headline points at the dramatic version, an agent that wanders off and does damage nobody asked for. There's evidence for that, too. OpenAI says it has notified over 100 organizations about activity by its own models.

The case an owner should think about is quieter. It's the agent doing a job you gave it, with access you gave it. That makes the map of what your agents can touch the same map you need to take to your broker.

In August, Reuters reported that cyber insurers including MSIG, QBE and Beazley were reviewing their policy language for autonomous AI. Its example was a company that gives an agent access to its network to fix security holes. "That could result in a loss, with no conventional hacker and potentially no unauthorized access at the outset," Reuters wrote. Insurance Business reported last week that cyber wordings which assume a human attacker risk leaving gaps.

General liability has its own version, and it has been around since January, when three optional generative AI exclusion forms took effect. Insurance Business, citing a nationwide review of state filings, counted 2,369 generative AI exclusion records in force in 49 states by mid-2026. Those are filing records, so they don't tell you how many policies carry an exclusion. The same review found no standalone product from an admitted, meaning state-licensed, insurer that puts the excluded coverage back.

Put those together and an agent that can send email from the company account, issue refunds, edit customer records, or delete files can cause a loss in each of those places without anyone breaking in. If you can't list those systems, your broker has nothing specific to check against your policy, and you find out what's covered when the claim comes back.

There's a fair case that this is less alarming than it sounds. QBE's view, as Reuters reported it, is that AI is "a risk amplifier" for existing cyber risk, and the insurer doesn't treat it as a new category. If that holds, plenty of existing cyber wording may respond to an agent's mistake much as it would to an employee's. The list is how you find out whether yours does.

So write it before the next renewal. For each agent, note what it can read and change, and who approved that access. It's the same discipline as fixing the work before buying the tool, applied to risk. RVRBN builds AI systems that let a small business run without its owner, so we have an interest in owners putting agents to work and in how that risk gets handled.

Then take the list to your broker and ask two questions. Does my general liability policy carry any of the three generative AI exclusions, CG 40 47, CG 40 48, or CG 35 08? How would my cyber policy treat a loss caused by an agent I authorized, where nobody broke in? Get the answers in writing. Insurance Business argues that the time to find exclusion language is before renewal, "while there is still room to negotiate."

Over the next few renewals, watch whether cyber carriers publish wording that speaks directly to losses an authorized agent causes. Beazley told Reuters it is developing new coverage. And watch whether an admitted insurer offers a product to fill the gap the general liability exclusions leave.

If you want to talk through what this means for your company, get in touch.