Tim Hardcastle (pictured), chief executive and co-founder of INSTANDA, said insurers need to take the latest warnings about AI seriously, but that reacting to each one is not the same as keeping pace with the risk.
Concerns about longer-term, systemic scenarios require a different response from the risks businesses are already encountering as they use AI, he said, and the market can act on those now.
That work starts with a clearer understanding of how businesses use AI and where existing policies already cover or exclude it, even when the wording doesn’t say so. Insurers also need to establish what effective oversight looks like. Underwriters are already considering those questions, and their understanding will develop with each claim and renewal.
The practical challenge is whether insurers can update their products as quickly as the risk changes. Changes to policy wording or rating models have often taken so long that the risk has moved on before the product catches up.
“Underwriters have the expertise to do this; the difficulty has often been their ability to put that knowledge into practice quickly enough to keep pace with the risk,” he said.
Hardcastle said the starting point is to identify specific events that can be insured rather than trying to cover AI as a whole. These could include a model producing an incorrect output that causes financial loss, a system breaching a contract or regulation, or an AI tool behaving unexpectedly or failing a required control. Risks like these are where the market is developing cover, he said, “and where I expect most actual claims to arise”.
Cover needs clear triggers and sensible limits, with pricing based on what an underwriter can assess, including how the system is used, who oversees it and the controls in place.
Insurance has always shaped behaviour as well as paid claims, he said, as when cyber insurers’ renewal requirements encouraged businesses to improve their security. By being clear about the controls they expect around AI, insurers could help establish standards for responsible use.
He acknowledged a real risk of gaps in cover but said blanket exclusions or unaffordable premiums will not resolve it. Insurers instead need products they can revise as evidence develops.
Unexpected behaviour is only part of the coverage question. As Insurance Business reported in September, Verisk’s Jenny Soubra sees exclusions emerging at the edges of cyber cover where an AI agent works exactly as designed but makes a costly decision independently. Those exclusions could narrow the cover available to clients.
Hardcastle drew a firm line at the extreme end of the warnings.
“If the concern is an AI system escaping anyone’s control and causing damage across the economy, I do not believe anyone can offer meaningful cover, and I would be cautious about claims that they can,” he said.
A systemic risk on that scale is more a matter for governments and capital markets than for individual insurance policies, he said. Anthropic’s IPO prospectus, filed this week, also raises questions about the limits of insurance. It flags “catastrophic or existential risks to humanity” if the technology is not developed with appropriate safeguards. It also warns that a catastrophic AI incident could exceed what any single insurer, and possibly the market as a whole, has the capacity or expertise to underwrite.
For the risks that can be insured, he pointed to cyber as a precedent. Twenty years ago that risk was difficult to define. Today it is a mature class with established policy wordings and claims data. He expects AI to follow a similar course, although faster.
Getting there will require insurers to turn underwriting decisions into product changes more quickly.
