As the number and cost of agent liability cases inevitably grows, insurance carriers are going to ask for evidence of constraints on agents in the form of rich policies, strong policy enforcement points, and collected logs that demonstrate your diligence or you will find that insurance becomes unaffordable or even unavailable.
Post 4 of a series on AI liability for general counsels interested in the legal liability associated with building and using agents.
Your cyber policy just got a new exclusion. Or a new questionnaire. Or a quiet note from the broker that agentic systems are now underwritten separately and the carrier wants to see controls before they will quote. The question is no longer whether insurers care about agents. The question is what they will accept as proof that you are not simply hoping the next model is safer (or as Satya Nadella calls it alignment).
Insurers writing AI liability coverage, or cyber policies that now reach agentic systems, want evidence, not promises. They have already seen the public incidents. They have seen agents ignore guardrails, chain exploits, and collaborate in ways that exceed any single permission. They know that a policy document and a vendor attestation do not tell them what an agent was actually allowed to do on a given Tuesday. They want the record that shows the ceilings were in force and that the runtime honored them.
The immediate effect is price. Companies that can produce an inventory of production agents, mapped controls, signed policy artifacts, and continuous traces will see better terms. Companies that cannot, will see higher premiums, higher deductibles, or lower coverage limits for agent-driven loss. The larger effect is access. Some risks become insurable only because the evidence exists. Without it, the carrier declines the exposure or carves it out. A tool that produces that evidence can pay for itself in premium reduction, or in the simple fact that coverage remains available.
Best practices still matter. Carriers will ask whether agents run in sandboxes, whether secrets are isolated, whether spending ceilings exist, whether human approval is required for irreversible actions, and whether you can show the effective permission was the intersection of the ceilings rather than the loosest one. They will also ask whether you red-team composed workflows and whether you can export the evidence when they ask. History of incidents and near-misses will be reviewed. The difference between a best-effort answer and a verifiable one is the difference between a renewal conversation and a coverage fight.
The practical requirements are the same ones that reduce the underlying risk. You need an inventory of production agents with clear ownership. You need published acceptance criteria that define data classes, tools, budgets, handoff rules, and human-approval points. You need those criteria compiled into a signed contract the runtime enforces. You need observability that records tool calls, policy decisions, handoffs, and approvals against that contract. You need the ability to hand an underwriter or claims adjuster the signed artifact plus the corresponding traces without reconstructing five consoles for each decision.
At Trustabl we treat the insurance evidence package as a direct output of the same controls that limit liability. Company Policy is the acceptance criteria the organization sets before an agent is allowed to run. The compiler turns it into a single signed contract and generates the corresponding policies for the Policy Enforcement Points (PEPs), so they stay in sync. The Safe Agent Control Plane records the actions, the decisions, and the approvals in one evidence bundle. The open-source Safe Agent Builder tools help teams produce agents that can actually meet those criteria. We do not claim this guarantees a lower premium. We claim it gives the carrier something verifiable to underwrite against, which is the difference between coverage that exists and coverage that does not.
Agents will keep creating new loss scenarios. Insurers will keep asking for proof that you constrained them. The companies that can produce the signed policy and the traces will be in a different position from the companies that can only describe their intentions.
Additional reading in this series
Related Blogs
Agent Reliability Resources
