Use an AI agent to review meeting action items

Define the review before adding automation

Pick one narrow job: tell explicit commitments apart from proposals in a single meeting. The agent doesn't need to send messages, create tasks or book anything. Keeping it focused means you can check every answer against the conversation.

“I can investigate that” and “I'll send the result Thursday” are different levels of commitment. A good review keeps that difference and flags what's missing. The agent's output is a suggestion to check, not an order that assigns someone work.

On any agent platform, first confirm it can read the transcript and that action tools it doesn't need are off. A prompt alone doesn't grant transcript access or permissions.

Set it up in Bridge

Custom agents need a paid Bridge plan. Open Create Agent from the sidebar or Account settings → Agents. You'll go through four steps: Identity, Abilities, Tools and Triggers.

Give the agent a clear name and instructions. Every agent can read transcripts, so there's no switch for that.

For a text review, it doesn't need hearing, screen or speaking abilities. Leave the optional tools off: web search, notes search, memory FAQs search and connectors. Rely on the @mention trigger, which is always on, and leave the other triggers off.

Save, then open the Assistant on the completed meeting and type the agent's @handle. The Assistant on a recap only sees that meeting's notes and transcript, and it can work with your agents.

Instructions you can adapt

Start with something like this, and test it on a meeting you can check yourself. It's a starting point, not a built-in template.

• Review the transcript available in this meeting context.
• List explicit commitments separately from tentative proposals and open questions.
• For each candidate action, give the task, stated owner, explicitly agreed date and supporting passage.
• If an owner or date is missing or unclear, write “not confirmed” rather than guessing.
• Don't contact participants or create tasks; produce a review for a person to check.

Test it on a passage you know

Say the transcript contains: “I can look into the export issue, but I need to check my workload. We could aim for Friday.” The reviewer shouldn't turn that into a confirmed Friday deadline. Here's what a useful review looks like (an example, not real agent output):

Instructions and tool access are different controls

Telling the agent not to send messages is a request, not a technical limit. The real control is which tools are switched on. Leave unneeded tools off, and check what's enabled before adding anything that can act outside the chat.

In Bridge, the @mention trigger is always on. The others are optional: proactively during meetings, by name, after every meeting and on a schedule. Tool access is a separate step. For a first test, turn on only what this job needs.

Check a few reviews before you extend it

For each action, read the cited passage and check who actually accepted it. Look for dropped conditions, wrong speakers and suggestions presented as decisions. Include a meeting with no agreed owner or date to see whether the agent invents one.

When you find a repeated problem, revise the instructions and rerun the same test. If you later automate follow-ups, check those actions and permissions separately. A good review doesn't mean automated follow-up is ready.