Before an AI agent
gets access.

An agent should start with a defined task and a clear owner. Broad access is not a substitute for a useful workflow.

Ask five questions

1. What is the outcome? Describe the result in a way an operator can check.
2. What data is needed? Limit access to the systems and records required for the task.
3. What can change? Separate information retrieval from permission to execute changes.
4. Who reviews the result? Define approvals, escalation and stop conditions before the workflow runs.
5. What proves it worked? Capture relevant test conditions and observed behavior, including limitations.

These questions are a planning aid. The final controls depend on the environment, data and consequences of the action.