Before an agent goes live, define its limits.

Separate assistance from authority

An agent that prepares a draft and an agent that sends it have different responsibilities. Describe exactly what the system can read, propose, change, and execute. Give each action an owner and an explicit boundary.

Design the handover to a person

Human oversight needs a place in the workflow. Decide what the reviewer sees, what evidence supports the recommendation, and what happens if the reviewer rejects it. Avoid relying on a final approval button without enough context to make a decision.

• Access: Use only the systems and information the task requires.
• Approval: Identify actions that need a person’s confirmation before execution.
• Testing: Include incomplete inputs, conflicting sources, and requests outside scope.
• Escalation: Make it clear when and how the agent should stop and ask for help.
• Recovery: Define how to pause the workflow and correct an unwanted change.
• Ownership: Assign responsibility for monitoring, maintenance, and reviewing changes.

Test the difficult cases

A successful demonstration shows what is possible. A useful evaluation also explores what happens when the conditions change. Review representative examples, record the failure modes, and agree which are acceptable before expanding access.

Measure the whole workflow

Track the time and effort needed for review, corrections, and follow-up alongside the output itself. Assess whether the process is better for the people who use it and those affected by its results.