AI for the Meeting Audit: Five Prompts That Do the Grunt Work

A Scrum Master I coach blocks two hours every other Tuesday to prepare for sprint review. She calls it her translation time. Down the hall, her calendar holds six other recurring blocks she is fairly sure should be a written update, but reworking them costs time she does not have. So they keep happening, live, every week, mostly out of inertia.

When teams reach for AI here, they usually point it at the wrong thing. They drop an AI notetaker into the meeting, get a tidy summary, and feel productive. A faster summary of a session that should not be live is still waste, just better documented.

Automating a meeting that should have been an email does not fix the meeting. It makes the theater run on time. If you read the piece on cutting review prep with AI, you saw the better version of this: AI earns its keep on the preparation, not on the performance.

This is the natural companion to this week's audit of your agile calendar. The audit decides what each session is for. AI helps you act on that decision in minutes instead of months of good intentions.

Make the judgment first, then hand AI the grunt work

Here are five workflows that use AI for the parts it is genuinely good at, the translating, the synthesizing, the structuring, while you keep the calls only a human should make. Each prompt is built for messy, real input, because that is what you actually have on a Tuesday afternoon.

1. Score a recurring session against the three questions

Paste whatever you have from the last occurrence. The prompt checks for a decision, a learning, and an owned action, then flags whether the session reads as conveyance (one-way information) or convergence (people deciding together).

I am auditing a recurring team session. I will paste whatever I have:
a transcript, an agenda, or rough notes. It will be messy and incomplete.

Do four things:
1. List every decision actually made. If none, say "No decisions found."
2. List every learning that will change what the team does next. If none, say so.
3. List every next action with a named owner. If an action has no owner, flag it.
4. Based only on what you see, tell me whether this reads mostly as conveyance
   (one-way information) or convergence (deciding together). If the input is too
   thin to tell, say "Not enough signal" and name what you would need.

Do not invent decisions, owners, or outcomes that are not in the text.
End with one sentence: keep live, move async, or redesign, and why.

Input:
[paste transcript, agenda, or notes]

One failure mode to watch: if you paste only an agenda, the model will happily invent decisions that never happened. Treat an agenda-only score as a hypothesis, and paste the real transcript when you can.

2. Convert a status meeting to async

Once you decide a block is pure information, AI can draft the written replacement so the live slot disappears cleanly.

I am replacing a recurring live status meeting with an async written update.
Draft two things:

1. An update template each person fills in before a deadline. Three prompts
   maximum, answerable in under two minutes.
2. A short recap the facilitator posts afterward that highlights only blockers
   and decisions needed, not a transcript.

Plain language, no jargon, no status theater. A busy person should read the
recap in 60 seconds. If you need to know our tools or cadence to make it
concrete, ask me one question first.

Context: [team, cadence, what the meeting currently covers]

If the model hands you a ten-field template, push back. Anything longer than three prompts will not get filled in, and the meeting quietly creeps back onto the calendar.

3. Build the pre-read for the sessions worth keeping

A convergence session earns being live, but only if people show up prepared. AI turns a messy doc into a one-page brief, so the least prepared person in the room is no longer the one who runs it.

Turn the material below into a one-page pre-read for a 30-minute decision
session. Structure it as: the decision we need to make (one sentence), the
three things you must know to weigh in, the options on the table, and the
open questions.

Cut anything that is context for its own sake. If a section has no real
content in my material, write "TBD, owner to fill" rather than inventing it.
Keep it under 400 words so people actually read it.

Material:
[paste the messy doc, thread, or notes]

Watch the background section. If it runs longer than the decision and the options combined, tell the model to cut it in half. The pre-read exists to start the conversation, not to replace it.

4. Redesign a decayed agenda around a decision

When you choose to keep a session but fix it, AI is good at rewriting the agenda so every block ends in something real. A sprint review, for instance, should be built around a decision, not a demo.

Here is the current agenda for our [event name]. It has drifted into a
[status report / demo / round the room]. Rewrite it so every block ends in
a decision, a learning, or a clear next action.

For each item, give me: the question that opens it, the timebox, and what
"done" looks like for that block. Replace any "updates" or "round the room"
items with either an async alternative or a real decision. Keep the whole
thing inside [X] minutes.

Current agenda:
[paste]

If it keeps a generic team updates block, that is the tell that it defaulted to the old format. Ask it to either cut that block or convert it to a written pre-read.

5. Turn the audit retro into commitments that stick

After you run the audit with your team, the board is a mess of stickies and side comments. AI is good at pulling the real commitments out of the noise, as long as you stop it from inventing owners.

Below are raw notes from a retrospective where we audited our recurring
sessions. Pull out only the commitments we actually made. For each, give me:
the change, the owner, the sprint it lands in, and a review date.

If a commitment has no owner or no date in the notes, flag it as "needs owner"
or "needs date" rather than guessing. Ignore venting, side conversations, and
anything that is not a committed change. Return a simple table I can paste
into our board.

Notes:
[paste]

The model will try to be helpful by assigning owners itself. Do not let it. An owner the team did not agree to is a commitment nobody made.

Where this goes wrong

The first trap is trusting the conveyance-versus-convergence label the model gives you. It has no idea what your team actually needs, so treat that call as a conversation starter, not a ruling. The harder version of this problem, when the tool says one thing and your instinct says another, is worth its own read.

The second trap is automating a session you should have cut. If a block fails the audit, the move is to remove it, not to give it a nicer summary. A clean recap of a dead meeting just helps it survive another quarter.

The third trap is feeding the model no context and expecting magic. Generic input gets generic output, so paste the real transcript, the real agenda, the real mess. And watch that a tidy summary never buries the owner or the decision, because a neat paragraph that loses who-owns-what is worse than messy notes that keep it.

Try this next week

Pick your most-dreaded recurring meeting. Grab whatever you have from the last one, a transcript, an agenda, even rough notes, and run workflow one. Read what comes back, then make the call yourself: keep it live, move it async, or redesign it.

You will have done in ten minutes what most teams avoid for years, and you will have evidence instead of a hunch. If your Scrum Masters want to build this kind of practical AI habit into how they run the team, it is the spine of our AI for Scrum Masters course.