Clients often ask us how much an agent should do on its own. The honest answer is: less than the demo suggests, and more than the risk team fears. We use three rules.

Rule one: reversible or not

If a step can be undone in a minute, an agent can take it alone. Tagging a ticket, drafting a reply, filling a field. If it cannot be undone, such as paying, sending to a client or deleting, a person approves it.

Rule two: show the working

Every decision an agent makes carries its evidence: the source page, the rule that fired, the confidence. A reviewer should be able to check a decision in under ten seconds. If they cannot, the interface is wrong, not the reviewer.

Rule three: sample forever

Even fully automated steps get sampled. A small share of every day's work is reviewed by a person and scored. The day the score drops, we know before the client does.

These rules make agents slower to demo and much faster to trust. In our experience the second matters more.