The steering committee of an AI project: who, how often, what agenda

What the committee is for

To decide — not to inform (the status report is read beforehand) and not to coordinate (that is the daily work between teams). Its decisions are the ones that exceed the team: move a date, cut scope, release budget, arbitrate between projects, accept a risk. When agents execute work, the committee gains one more job: checking that the agents acted within the frame they were given.

Five people, not fifteen

The sponsor, the project lead, the leads of the teams involved, the person who governs the AI agents (roles, permissions, autonomy modes) and — unless a cockpit records decisions as objects during the meeting — someone to write them down. Beyond six or seven people the committee informs instead of deciding.

Weekly and short beats monthly and long

A two-hour monthly review always comes too late: gaps are three weeks old and pending decisions have piled up. Thirty minutes a week on a prepared briefing settles things while the gap is still small. With active agents, a weekly review of their actions is the minimum.

The agenda: three questions, then a fourth

1. Are we on trajectory? State per goal, margin consumed, what moved since last week.
2. What threatens the date or the scope? Signals off trajectory, risks whose probability changed, slipping dependencies — each with a proposed response.
3. Which decisions are pending? Each with its options and their cost. Decide, name an owner, date it.
4. What did the agents do? Direct actions, proposals, refused proposals and why. This is how permissions get tuned: an agent whose proposal is approved ten times can go direct on that action; an agent whose proposals are refused is badly framed.

Then: last week’s decisions and their effect.

The briefing an agent prepares

The concrete change: the briefing no longer costs half a day of compilation. The agent reads goals, remaining work, capacity, blockers, open decisions and risks, and drafts the state per goal, the signals off trajectory and the options for each pending decision. Three rules keep it useful: it cites its sources, it separates facts from proposals, and it never decides — in Stellary the briefing agent runs in supervised mode.

FAQ

Who should sit on the steering committee of an AI project?

The sponsor, the project lead, the leads of the teams involved and, when AI agents execute work, the person who governs their roles and permissions. Beyond six or seven people the committee informs instead of deciding.

What does an AI agent change in a steering committee?

It prepares the briefing from the project data, with sources, so the committee arrives to decide rather than to understand. And it adds a question to the agenda: what did the agents do on their own, on proposal, and what was refused?

Keep reading

• Project management vs project piloting
• The cockpit guide
• AI agent orchestration