Trust & Evidence

How governed AI agents work

Velerian's agents do real claims work — and show their work. Here's what that means, without the jargon.

An agent is a worker, not a chatbot

A chatbot answers questions. An agent does a job: it takes in a task — say, checking whether a policy covers a particular loss — works through the steps, and produces a result you can act on. The difference that matters: an agent has a defined job and defined limits, so you always know what it is and isn't allowed to do.

"Governed" means it works inside rules it can't break

Every Velerian agent runs inside guardrails. Some are about authority (what the agent may decide on its own versus what it must escalate), some are policy checks (it has to reference your actual policy language and claims-handling rules), and some are hard rules it can never break — for example, it will never issue a denial without citing the specific policy clause behind it. The agent doesn't get to improvise around these.

Nothing ships until it passes a check

Before any result is released, it goes through an automatic reviewer — an evaluator — that tests whether the work is sound and properly supported. If it doesn't pass, it doesn't go out. And if the agent isn't confident or can't back up its answer, it stops rather than guessing. That last part has a name — fail-closed — and it's the opposite of how most AI tools behave.

A person still makes the call

The agent does the heavy lifting; a qualified person makes the final decision. Your adjuster or examiner stays accountable — the agent hands them a finished, documented recommendation to approve, not a decision made behind their back.

Every decision comes with its evidence

This is the part regulators and auditors care about. Each decision arrives with the reasoning behind it, the sources it cited, and a record of who reviewed it — its evidence trail. When someone later asks "who decided this, and on what basis?", the answer is already attached.

The shape of every decision

  1. A task comes in — with the policy, the facts, the file.
  2. The agent works it, inside its guardrails.
  3. An automatic evaluator checks the result; weak or unsupported work is stopped.
  4. A qualified person reviews and signs off.
  5. The decision goes out with its evidence trail attached.

New to any of these terms? See the Glossary.