Give AI Agents an Authority Budget

Give an AI agent neither a blank cheque nor a human checkpoint for every step. Define its authority before the workflow begins:

• Proceed: reversible actions within an agreed scope, such as sorting requests or drafting an internal summary.
• Prepare, then ask: actions that affect another person, spend resources, change access or create an external commitment.
• Stop and escalate: actions outside its permissions, or when essential context is missing.

Name the human decision owner, the evidence the agent should present, and what happens if approval is delayed. Then review a sample of actions and near-misses; revise the boundary when the evidence calls for it.

The trade-off is speed versus the cost of a mistaken action. Make that choice explicit—and keep responsibility with the person who owns the decision.