Most teams that give an AI agent real access spend their effort on what it can do. Fewer sit down and sort those actions by one question: can this be undone?
The difference is easy to see with email. A draft can be edited, reread, or thrown away. A sent email can't be unsent. If an agent drafts a wrong message, you lose a minute. If it sends one, a customer has already read it.
Two kinds of actions
Cheap to get wrong (reversible):
- Drafting a reply or a document
- Searching and summarizing
- Proposing a change for someone to review
Costly to get wrong (hard or impossible to undo):
- Sending messages to customers or partners
- Deleting files or records
- Making payments or issuing refunds
- Publishing something publicly
- Changing a customer record
Many setups hand the agent both kinds of permission in one lump, so a vague request and a confident guess can turn straight into an irreversible action.
Make a list, then add a person
The practice is simple. List every action your agent can take. Mark the ones that can't be undone. Then require a person to approve each of those before it runs.
A few details make it hold up:
- Enforce it in the tool, not just the prompt. If approval only lives in the instructions, a model can talk itself past it. If the tool can't run without a recorded approval, it can't.
- Show the approver what will happen. "Approve action?" invites clicking through. "Send this exact message to this customer?" is a real decision.
- Revisit the list when you add a tool. New integrations quietly add new irreversible actions.
Questions for you
- Do you keep an explicit list of irreversible actions, or decide case by case?
- Which action turned out to be harder to undo than you expected?
- How do you keep approvals from becoming a rubber stamp?
Disclosure: I run a small AI agent risk-check service called AgentRisk, which is why this topic is on my mind. More at https://rleffke.com/agentrisk
Top comments (0)