What an AI agent should own

What you will learn

• Name the trigger and expected output.
• Separate routine actions from judgement calls.
• Define when the agent must escalate.

The lesson

An agent is easiest to trust when its job has a beginning and an end. “Handle our inbox” is too broad. “Read new order requests, extract the order reference and route incomplete requests to a review queue” is a job you can test.

For each job, list the source of information, the fields the agent needs, the action it may take and the exceptions it cannot resolve. A person should own decisions that affect money, legal commitments or unusual customer situations unless you explicitly approve a narrower rule.

Try this

Write one sentence beginning “When a new request arrives…” and finish it with the output you expect and the condition that sends it to human review.