Where should AI agents act autonomously, and where should people stay in control?

Classify actions by consequence

The useful unit of design is the action, not the agent. List what the system can do, then assign each action a level of authority. A single workflow may need several levels at once.

• Read: retrieve approved information without changing a record.
• Recommend: draft a decision or next step for a person to review.
• Approve: require a named person to authorize a consequential action.
• Execute: allow the agent to act within a defined limit, with an audit trail and a reversal path.

Define ownership and escalation

Every action needs an accountable owner, a reviewer where consequence requires one, and a place to escalate an exception. Write these responsibilities into the workflow rather than leaving them to the agent prompt.

Collect evidence before expanding autonomy

Run the agent on representative cases and record completion, corrections, tool selection, failures and recovery. Expand authority only when the evidence shows the workflow can absorb the consequence of a wrong action.