An AI agent to run our meetings

Nora is an AI agent I built to run our follow-up meetings at Oliverlist and publish the minutes. She builds the agenda from the numbers, takes the items in turn and records the decisions. She has only just started, so there are no results to share yet. Here is why I built her, what I expect for decision follow-up, and what I will not hand over.

Why hand over the running of a meeting

Follow-up meetings tend to fail in the same ways. Part of the time goes into finding the figures. Decisions from last time are half-forgotten. The minutes arrive late, or never, because nobody has time to write them. This is not about goodwill. Taking notes, chasing and checking are invisible tasks, so nobody does them.

My follow-up meetings already start from a solid base. Every night, diagnostics check Oliverlist's figures and publish reports in which only anomalies are escalated. What was missing was someone to turn those reports into an agenda, hold the thread during the meeting and write down what was decided.

That role usually falls to whoever chairs the meeting. It is a poor combination: you cannot run the meeting, take notes and make the decisions at the same time. And since part of the team works remotely, these meetings happen on video anyway. Having an agent join the call is a small step from there.

What Nora does

Before the meeting, Nora builds the agenda from the reports. Everyone can read it in advance. A good follow-up agenda has few items: the week's anomalies, the decisions still pending, the figures that moved. Each item ends with a question to settle. Spoken aloud, a question that does not fit in one sentence does not get settled, so the questions have to stay short.

During the meeting, she takes the items one by one, presents the figures, asks the questions that need a decision and records the decisions. She speaks, but in front of people who can correct her at any moment. That is the real difference from an agent calling a prospect, as I explained about our voice agent on hold: here, a mistake gets fixed in the room.

After the meeting, the minutes are published without anyone having to write them.

The cost is a few dollars per meeting. Set that against the combined time of everyone around the table, and against the time spent writing the minutes, or not writing them. The real cost lies elsewhere: the time it takes to get her right, then to read her first sets of minutes closely.

What changes for decision follow-up

A decision that is not written down does not exist. That is the first thing Nora changes: every decision is recorded, and the minutes list them.

I also designed her to go further than note-taking. When a decision concerns a check or another agent, such as adding a test to the nightly diagnostics, it can be applied straight away. Three safeguards frame it: a dated copy of the previous version, a record of who asked for the change, and a list of changes in the minutes.

A remark that is not a decision does not get lost either. It goes onto a list of proposals to settle later.

The goal is simple: every meeting picks up from what was decided last time, and nothing depends on anyone's memory. It is also a matter of credibility for a CEO. A decision taken in a meeting and never applied teaches the team that meetings are pointless.

What stays human

Nora runs the meeting; she does not decide. Decisions stay with the people around the table, by design.

I also keep everything about people outside her scope: individual feedback, disagreements, performance reviews, difficult announcements. Those topics need tact and trust, not an agenda.

Introducing the agent is a human job too. Brief the participants before the first session: who she is, what she does, what she does not do, and how to interrupt her. A meeting run by a machine is surprising. Better for the surprise to happen before the session than during it.

Finally, an agent running a meeting needs watching on two points. She must quote dated figures: last week's number presented as today's is enough to skew a decision. And her notes do not replace the recording: when there is doubt about who said what, the recording is the reference.

Key takeaways

• A follow-up meeting run by an agent starts from the numbers, not from memory.
• What is at stake is the follow-up: decisions that are written down, applied and traceable.
• Put limits on what the agent can change: a copy before every change, the requester's name, a list in the minutes.
• Keep decisions and people topics with people.
• Brief participants before the first session: what the agent does, what it does not do, and how to interrupt it.
• Wait for several meetings before judging. A first session measures nothing.