You can't insure what you can't describe

An AI agent is not a chatbot. It books, buys, files, transfers, and reconfigures. It takes actions in systems that matter. The moment software acts on its own authority, the question stops being "was the answer good?" and becomes "was that action allowed, and who pays if it wasn't?"

The underwriting problem

Insurance needs a describable risk. For an agent, that means answering three questions before an incident, not after:

1. Authority: what is the agent actually permitted to do?
2. Enforcement: what stops it from exceeding that authority?
3. Evidence: how would anyone know if it did?

Most agents today can't answer these crisply. That isn't a coverage problem; it's a legibility problem. You can't price what you can't describe.

Certification before coverage

Certification is our answer: an evidence-based accreditation of an AI agent's authority boundaries, controls, and oversight, in the spirit of a SOC 2 for autonomous software. It's currently in alpha.

The sequence is deliberate: certify first, insure second. Certification establishes the common language for what an agent is and how it's controlled. Agentic E&O, errors & omissions coverage for deployed agents, is built on top of that substrate.

If you're deploying agents and want to get ahead of this, the certification program is where to start.