Safety
What should you never give an AI agent permission to do?
A practical permissions checklist for keeping personal and business AI agents useful, reviewable and safe.
6 min read
An agent should receive only the access required for its defined job. Convenience is not a reason to grant unrestricted account access, especially when actions are expensive, sensitive or difficult to reverse.
Keep high-impact actions behind approval
- Payments, purchases and financial transfers.
- Deleting records or changing permissions.
- Sending sensitive or legally significant messages.
- Medical, legal, employment or safety decisions.
Use least privilege
Prefer read-only access where possible. Separate accounts, limit the available tools and review access when the workflow changes. Record what the agent did so a person can inspect the result.
Test the exceptions
A safe workflow is tested with missing data, an incorrect request, an urgent request, an opt-out and a request outside the agent's role. The expected response should be a clear question or handoff.