What should a human review before an AI agent's work is used?
Before you trust an AI agent's work, decide what you need to see. For a meeting brief, that means where each fact came from and which items are still open. Write that down before the AI agent starts, and human review stops being a vague request.
Say an AI agent prepares a meeting brief. A clean layout is part of what makes it ready to use. The names, dates, sources and open questions need your eyes too.
A draft is a safe place to start
For a first build, a draft brief for you to review makes a sound boundary. The AI agent doesn't contact the client or change the meeting, and it doesn't distribute the brief. Connecting a tool and giving permission to act are separate decisions.
Then make the review specific. Ask the AI agent to show where each important fact came from and to separate what's confirmed from what's still open. When two sources disagree, it should keep both visible and leave the choice to you. Say the client's public website lists one job title and a recent message lists another. That's one to check before an introduction, and a polished paragraph would bury it.
"It drafts. You decide what goes out" works as a starting boundary once you've spelled out what you need to see first.
What to try this week
Test the instructions on a task you can check yourself. Read the draft against the original sources, and each time you correct something, add the rule behind the correction to the instructions.