The policies were written before the software could act on its own
An autonomous agent with system access can create a liability in seconds. The cover a company already holds was drafted for a world where a human made that call. Click to read why that gap matters.
Most commercial policies allocate liability by asking who acted and whether they acted reasonably. Professional indemnity asks whether a professional fell below a standard of care. Cyber asks whether an outsider got in. An AI agent that has been given write access to a company's systems and takes a wrong action fits neither question cleanly. Nobody broke in, and no professional was negligent in the usual sense. The company is still liable for what its systems did.
The regulatory direction is not ambiguous. The EU's AI Act places graded obligations on deploying organisations, not only on the companies that build the models, and those obligations phase in over the period most enterprises are actually rolling agents out. India's data protection regime puts the accountability for automated processing on the entity that decides how it is used. In both cases the buck stops with the deployer, which is precisely who currently cannot buy a policy priced for the exposure.
The pricing problem is narrower than it looks. It is not that AI risk is unknowable. It is that the three inputs an underwriter needs are all thin at once: how often an agent causes a loss, how large that loss runs, and how strongly incidents inside one organisation correlate with each other. Incident counts are sparse and inconsistently disclosed. Settled claims that would anchor severity barely exist yet. And correlation is the one that should worry a portfolio, because a fleet of agents running the same model on the same prompt scaffolding does not fail independently.
What we are doing about it: pricing this from the deployment's actual shape rather than waiting for a decade of claims. What the agents can reach, what a wrong action costs at each of those reach points, what human review still sits in the path, and how much of the fleet shares a single point of failure. Frequency and severity modelled separately, thin history blended against a class average, and an explicit statement of how much of the number is observed versus reasoned.
The honest caveat: a meaningful part of any AI liability score today rests on structural reasoning rather than fitted loss history, because the history does not exist yet. We would rather put that on the face of the assessment than quietly price it in.