Use an AI agent to spot project risks in meeting transcripts

The job: catch what was said in passing

Someone says "if the vendor slips, we miss the launch", and the meeting moves on. Nobody writes it down, because it wasn't on the agenda.

This agent does one narrow job. After a project meeting, it lists the risks people mentioned, and proposes register entries. It doesn't rate them for you or decide what to do.

Inputs, output and what it must not do

Part	What it should be
Input	The meeting's transcript and recap
Looks for	Statements that something could go wrong, with the condition and the effect
Output	A draft entry per risk: description, quote, speaker, suggested owner, open question
Stop rule	No risks found means it says so and stops
Asked not to	Invent a risk, assign a final owner, or change anything already in the register

Instructions you can adapt

This is a starting point, not a built-in template. Replace "Project Atlas" with your project's name, and test it on meetings you remember well.

If this meeting wasn't about Project Atlas, reply "Not a project meeting" and stop.
Find statements that something could delay, harm or block the project. If there are none, reply "No risks found" and stop.
Skip problems that have already happened. List those separately as issues.
Write each risk as: if this happens, then this is the effect.
Add the exact quote and who said it. If you have no quote, leave the risk out.
Suggest an owner only if someone volunteered or was named. Otherwise write "Owner needed".
End each entry with one open question.
Never invent a risk or assign a final owner. Don't edit existing pages. Create one new page titled "Risks to review" with the meeting date.

Set it up in Bridge

Create the agent in four steps: Identity, Abilities, Tools and Triggers. Paste the instructions in Identity, up to 4,000 characters. Reading transcripts is always on.

In Tools, turn on Notion so the agent can create the review page. In Triggers, choose After every meeting. Notion connections, custom agents and agentic workflows need a paid plan.

Keep a person in charge

The trigger runs for every meeting, which is why the first instruction is a stop rule. Name one person to check the output. Have them look at where the first run's page lands, since a run with no risks creates no page.

That person opens each quote in the transcript, and moves accepted risks into the real register.

"Asked not to" is a request, not a control. With Notion on, the agent can act on any page the connection reaches, up to everything you can access, so check its first few runs closely.

Prefer to see the list before anything is created? Turn Notion off and leave After every meeting off. Change the last instruction to "List the risks in this chat", then @mention the agent in the Assistant on a completed meeting.

Example output

An invented example, from a weekly project call:

Risk: if the vendor's API review slips past the 12th, then the launch date moves.
Quote: "If they don't sign off by the 12th, we can't launch that week."
Said by: the engineering lead.
Owner: owner needed.
Question: who is the contact at the vendor?

Limits to plan for

It only sees what was said. A risk nobody voiced won't appear.
It reads one meeting at a time, so it may repeat a risk already in your register. The reviewer removes duplicates.
It can misread a joke or a hypothetical as a risk, which is why every entry needs a quote.
It can't judge likelihood or impact from one meeting, so leave scoring to your team.
Each meeting needs a transcript, so calls you didn't capture are missed.

Sources & further reading

Keep readingUse an AI agent to review meeting action items →