Customised AI tools blur who is liable, while changing uses can leave cover lagging behind
Tim Johnson (pictured), partner and head of the insurance practice at Browne Jacobson, is seeing few disputes over AI so far. Businesses are cautiously weighing where liability might sit, he said, while lawyers have already faced action around the world for citing case names invented by AI.
“I think that’s where it’s starting, but you can be absolutely sure that civil liability claims will follow,” he said.
His colleague Philip James, partner and head of UK data, privacy and cybersecurity group, sees the potential for those disputes to be significant, because emerging technologies commonly bring with them changing responsibilities and are more complex and unfamiliar than well-established software applications.
Cyber disputes are another matter, Johnson said the gap between expectation and reality was the most common source of disputes between insurers and policyholders. A cyber policy can bundle covers from forensic IT to cyber extortion, but a business may not have bought the section an incident calls for. A firm whose operations are disrupted may find it never bought a cyber business interruption extension.
That is not necessarily a flaw in the product, business owners can treat buying a cyber policy as completing their insurance checklist.
“And they think, okay, cool, got cyber policy, tick,” he said.
The risk with cyber, Johnson said, is that “it’s potentially too easy to buy because people aren’t necessarily taking the time or taking the advice to make sure it matches up.”
Who is the supplier?
With AI, the harder question is who is liable when something goes wrong. Parts of EU law have drawn on old product liability legislation to separate AI developers from deployers, but the line is rarely clear-cut because businesses customise what they buy and build on top of it.
“And then it’s like, oh, well, we thought we were the deployer, but now we’re maybe part deployer and part supplier,” he said.
Agentic AI complicates that further, James said. Beyond the legal question of how far an AI agent’s authority extends, even vendors themselves have expressed a desire to slow development until the technology can be implemented, and its full capabilities harnessed, securely.
Johnson compared that customisation with a vendor mixing, treating or adapting a manufacturer’s product, which he said would typically fall outside the manufacturer’s vendor cover. AI adds another complication.
“But I think with AI, it’s even more complex because the product itself is doing some of the changing to itself,” he said. That change comes partly from users’ prompts and partly from the model’s own characteristics, and no one can look inside the box.
What brokers ask
Johnson cautioned against reducing a client’s AI exposure to a single renewal question.
“The one question I wouldn’t ask as a broker is, are you using AI? We’re beyond that now,” he said. AI is built into everyday software such as Outlook, so it is almost impossible for a business not to use it in some form.
Brokers who understand the AI tools used in their clients’ markets can probe further, Johnson said. They ask what guardrails are in place, who is feeding in the data, who reviews the outputs and whether the tools are patched.
The exposures behind those questions are already live, either in the public domain or in matters the firm has seen. They include staff putting confidential information into free, non-enterprise tools, vendors training models on client data, sometimes framed in contracts as service improvements, and recording tools used without everyone’s consent.
After speaking with large US brokers at a recent conference, James said he could see specialists working alongside insurers to help clients check their controls, particularly smaller customers. He compared it with checking the locks and alarms on a house, though cyber calls for much closer work with technical security experts.
“I can certainly see a market where there’s going to be more time spent under the bonnet,” he said.
Cover that keeps pace
Johnson said a business’s AI use 12 months from now may look very different from today, so brokers should remind clients to tell them about changes during the policy year. They might also check in when mainstream software adds a new AI function.
“If something’s changed midway through the policy year, you should proceed on the assumption you’re not going to be covered for it unless you told your insurers,” he said. “That’s the cautious way to proceed.”
Johnson’s anecdotal impression is that AI risks have become more apparent over the past year. Some businesses that initially rushed in are slowing down, having seen that it “can also do bad stuff very quickly and very well”.
That more measured approach gives brokers a chance to find out what clients are actually doing with AI before a claim tests what their cover was built for.
