blast radius: the only number that matterswhen insuring an agent

Ask a room of engineers about an AI agent and the first question is almost always about the model. Which one, how big, how new. Ask an underwriter the same question and the model barely registers. What we want to know is different and more revealing: what can this agent do, unattended, before a person can stop it? That number is the blast radius, and it is the single best predictor of a loss we have found.

why the model is a distraction

The same model is a rounding error in one deployment and a catastrophe in another. Wired to a read-only knowledge base, it can embarrass you at worst. Wired to issue refunds, move money, change infrastructure, or delete data, it can end a quarter. The model did not change between those two cases. The authority did. Pricing a risk from the model name is like pricing a driver from the color of the car.

what actually sizes the blast radius

• i.financial authority: the largest amount the agent can move, refund, or spend without a human approving it. this sets the ceiling on a single bad event.
• ii.irreversibility: can its actions be undone? a wrongful refund that can be clawed back is a smaller risk than a deleted production database that cannot. reversibility collapses severity.
• iii.reach: how many customers, records, or systems can a single action touch? reach is where aggregation risk hides.
• iv.rate: how many actions can it take in a window before anything trips? without a circuit breaker, a single fault repeats thousands of times before anyone notices.

the controls that shrink it

Blast radius is not fixed. It is designed, and it can be reduced. A per-transaction cap enforced in code, not in a prompt. A daily aggregate circuit breaker. Human approval above a defined stakes threshold. Rollback and a kill-switch on anything irreversible. Each of these directly lowers either the probability or the severity of a claim, which is why each of them lowers your premium.

“in practice, the autonomy and blast-radius factors swing a premium by more than the entire rest of the rating sheet combined.”

Because we underwrite the blast radius rather than the model, you have direct control over your rate. Narrow the authority and the price falls, visibly. Two companies running the identical model can pay very different premiums, and the difference is not arbitrary. It is the difference in how much damage each has allowed its agent to do before a human steps in.

The industry will keep asking which model your agent runs. We will keep asking what it is allowed to do. One of those questions predicts a claim. Build your agent, and read your policy, accordingly.