When an AI agent should ask a person first
20 November 2026
An agent that can act on your systems needs a clear line between what it may do alone and what waits for approval. Here is how we draw that line.
An AI agent is useful because it acts: it looks things up, updates records and sends messages without someone doing each step by hand. The same ability is what makes it risky. The design question is not whether the agent can take an action, but which actions it may take without a person agreeing first.
Sort actions by what it costs to be wrong
We list every action the agent could take and ask one question of each: if this is wrong, how hard is it to undo? Reading an order, drafting a reply or flagging a mismatch is cheap to get wrong. Issuing a refund, changing a ledger or emailing a customer is not.
Actions that are cheap to undo run on their own. Actions that are expensive to undo stop and wait. Most workflows have a middle group, and that is where a limit helps: a refund under a set amount goes through, a larger one is prepared and held.
Make the handoff worth receiving
An approval step fails if it only says "please review". The person approving should receive what the agent found, what it proposes to do and the evidence behind it, so the decision takes seconds and not a fresh investigation.
- What was asked, in the requester's words.
- What the agent checked, and what each check returned.
- The action it proposes, and why that one.
Record everything, including what was approved
Every step is logged: the request, each tool call, the result and who approved what. The log is how you find out why something happened three weeks later, and it is how the limits get adjusted with evidence instead of guesswork.
The line between "alone" and "ask first" is not fixed. It starts conservative and moves as the record shows which actions the agent gets right.