Making an AI agent safe: human approval and an audit log

An AI agent is software that makes its own decisions and works with APIs and databases. The upside is big, so is the risk: if the agent cancels the wrong order or quotes the wrong price, who is accountable? The answer is architecture. Three things make an agent safe.

1. An action list with levels

• Write down every action the agent can take: "view order", "change status", "refund".
• Give each a level: autonomous (read, report), with approval (status, discount), forbidden (money, delete).

2. Human in the loop

• On "with approval" actions the agent stops and sends the manager a request in Telegram/CRM: what it wants to do, why, based on which data.
• Approval is one button; on rejection the agent takes the reason and proposes another path.

3. An activity log

• Every step is recorded: input, decision, API called, result, time.
• The log lives where it cannot be edited and is reviewed weekly — that is how you see the agent's habits.

Practical advice

Make the first agent read-only: it answers questions and builds reports but changes nothing. Watch the log for 2–4 weeks. Few mistakes — unlock the "with approval" level. "Forbidden" actions are never given to the agent; that is a design decision, not a technical limit.