When an AI agent should hand back to a human

The question that decides whether an AI agent is useful is not how often it is right. It is what happens the rest of the time.

1. Escalate on irreversibility, not on confidence

A model can be confidently wrong, and the cost of being wrong is not evenly distributed. Anything the business cannot undo needs a human confirmation regardless of how sure the agent is.

2. Escalate on novelty

When an agent encounters a case that does not resemble anything in its evaluation set, that is the moment to hand back — not because it will necessarily fail, but because nobody has checked whether it succeeds.

3. Escalate on repetition

If an agent has tried the same tool three times and is still not done, it is not going to be done on the fourth. Cap the attempts and hand back with what was tried.

4. Escalate when the human asks

Every agent-facing surface needs a visible, always-available route to a person that does not require the user to argue with the agent first.

The one that catches the most

Across the systems we have shipped, novelty catches more genuine problems than the other three together — and it is the rule most teams skip, because it requires maintaining an evaluation set rather than writing a threshold.

• #AI agents
• #escalation
• #production AI