How to use AI for meetings, notes, and follow-ups

AI is useful for meetings when you use it to turn what was said into what happens next. The value is not in storing a transcript; it's in extracting the decisions, the actions, and the open questions, and then closing the loop on them. Most people skip that step and end up with a tab full of unread "meeting summaries."

This guide is for the meetings you have every week: customer calls, internal syncs, 1:1s, interviews, project reviews. The core idea is simple — let AI handle the capture so you can focus on the conversation, then use AI again to convert the raw notes into something that moves work forward.

A 3-part working model

Most people only use AI for the middle step. The compounding value is in the bookends.

1. Before the meeting

Pull prior context. If there are previous notes, summaries, or threads on this topic, run them through AI: "Summarize what we agreed last time. What was left open? What changed since?"
Generate a brief. For an external call: "Here's the LinkedIn profile and recent posts of the person I'm meeting. Give me 3 talking points and 1 question they're likely to ask."
Surface questions. If you're walking into an unfamiliar topic: "I'm about to talk to a [role] about [topic]. What 3 questions would a sharp person ask in this meeting?"

The point is to arrive prepared without spending 30 minutes prepping. A 3-minute briefing is usually enough.

2. During the meeting

Use an AI note-taker. Granola, Otter, Fathom, Fireflies, or built-in meeting tools. They capture the transcript and a draft summary automatically.
Stay in the conversation. The point of the AI note-taker is that you don't take notes manually. If you find yourself typing during the call, you've lost the value.
Disclose if recording. Most jurisdictions and most workplace norms require consent. "I'm using a note-taker — okay with you?" at the start, every time.

3. After the meeting

This is where most people stop too early. The raw AI summary is a starting point, not the output.

Extract decisions — what was decided, separately from what was discussed:

From this transcript, list only the explicit decisions made. Not topics discussed — decisions.

Extract action items with owners — the single most underused move:

List every action item. For each: who owns it, what's the deliverable, what's the due date. If owner or deadline is missing, flag it.

Surface open questions — things that came up but didn't resolve:

What questions came up in this meeting that didn't get answered? Group them by topic.

Draft the follow-up — using all three above:

Draft a 6-line follow-up email to [participant]. Open with the one decision that matters most to them. Then their action items. Then any open question I owe them an answer on.

The follow-up email lands in their inbox 20 minutes after the meeting. That alone changes how meetings translate into work.

A real example

Take a 45-minute customer discovery call. The AI note-taker produces a 4-page summary. Almost nobody will read 4 pages.

Bad output (what most people accept):

The customer discussed their current setup, mentioned challenges with their existing vendor, expressed interest in our product, and asked about pricing. Next steps were discussed.

That's noise. It tells you nothing actionable.

Good output (what to ask AI to produce):

Decisions made:
- Customer will run a 14-day pilot starting Nov 18.
- Pilot scope: 2 of their 5 sales reps.

Action items:
- (Me) Send pilot contract by Wed Nov 13.
- (Me) Set up their workspace + onboard the 2 reps by Mon Nov 18.
- (Customer) Confirm the 2 reps + share their territories before Nov 18.

Open questions:
- They asked whether reports can be filtered by territory — I said yes but need to verify.
- They asked about EU data residency — I said I'd check; their head office is in Germany.

Likely red flag:
- Their current vendor contract renews in March. They're shopping seriously now because that's the budget moment. We need decision by mid-Jan, not after the pilot.

That summary you can act on. You know what to send, who needs to confirm, what to verify, and where the deadline pressure sits.

What most people get wrong

The patterns repeat across teams:

They save everything and use nothing. A library of transcripts is not a memory system; it's a graveyard.
They accept "meeting summary" as the deliverable. A summary describes the meeting. An action list moves work.
They don't separate decisions from discussion. A topic being raised is not the same as it being decided. AI summaries often blur the two.
Action items without owners and deadlines. "We should send them the pilot doc" is not an action item. "[Name] sends pilot doc by Wed" is.
They write the follow-up from memory the next day. Memory has already softened. The follow-up should go out within an hour, drafted from the actual transcript.

A good post-meeting prompt

Use this verbatim, adjust to your context:

Below is a transcript of a [meeting type] between [participants].

Produce:

1. DECISIONS — explicit decisions made, not topics discussed. If there were no decisions, say so.
2. ACTION ITEMS — for each: owner, deliverable, due date. If owner or date is missing, flag with [?].
3. OPEN QUESTIONS — things raised that didn't resolve, grouped by topic.
4. RISKS OR FLAGS — anything I should worry about. Time pressure, mismatched expectations, scope drift, churn signals.
5. SUGGESTED FOLLOW-UP — a 6-line email I can send to [primary participant] within an hour.

Skip:
- Restatement of what was discussed.
- "Strong engagement" and other generic vibes.
- Anything I can't act on tomorrow.

Transcript:
[paste]

This is the prompt that turns AI from note-taker to operations assistant.

Where AI helps with meetings

Capture without splitting attention. You can listen and respond instead of typing.
Decision and action extraction. Faster and more reliable than human note-takers for structured output.
Pattern recognition across meetings. "I've talked to 12 prospects this quarter. What are the 3 most common objections in the transcripts?"
Tone analysis at the edge. Spotting when a customer's language shifted from enthusiastic to cautious mid-call — a model reading the transcript can flag this faster than you can.
Follow-up speed. The 20-minute follow-up is achievable with AI. The 24-hour follow-up isn't.

Where AI goes wrong with meetings

It pretends the summary is the work. Generating a summary feels productive; the work is what you do with it.
It blurs decisions and discussion. Models lean toward "balanced summary" rather than crisp decision extraction. Force the structure.
It misses what wasn't said. A customer carefully not committing to something is signal. AI rarely flags omissions.
Names and titles drift in long transcripts. When the transcript misattributes a quote, the summary inherits the error. Spot-check.
Sensitive context leaks. If you record a 1:1 with a direct report and run it through cloud AI, that's now in a vendor's logs. Check your tool's data policy before recording anything sensitive.

When AI note-taking is risky or wrong

Confidential conversations. Performance, compensation, legal, M&A, anything HR-adjacent. Don't record.
Therapy-adjacent conversations. 1:1s where someone is unloading personal stuff. The recording itself changes the dynamic.
When participants don't know. Even if it's legal in your jurisdiction, recording without disclosure breaks trust. Always announce.
Regulated industries. Healthcare, finance, legal — your industry may prohibit certain AI services from touching meeting content. Check before, not after.
Customer calls where the customer says no. They have the right. Take notes by hand or assign a note-taker.

Common mistakes

Saving the transcript instead of the actions. A transcript is raw material, not output. Extract before archiving.
No owner, no due date. Action items without these are wishes, not commitments.
Waiting until tomorrow to send the follow-up. Speed is the point. Same hour, while the meeting context is intact.
Letting the AI summarize without structure. "Summarize this meeting" gives you what AI thinks a summary looks like. Tell it the structure you want.
Using one prompt for every meeting type. A customer discovery call and an internal sync have different output needs. Adjust accordingly.
Disclosure failure. Recording without telling participants damages trust faster than any productivity gain recovers it.

The summary, plainly

The value of AI in meetings is not capture. It's conversion — turning what was said into what happens next. Capture is the table stakes. Decision extraction, action items with owners, open questions, and a same-hour follow-up are the actual output. Treat the transcript as raw material, not the work.

Related:

Replace your weekly report with AI — same conversion pattern applied to weekly status
Calendar and meeting prep: never walk in cold again — the prep side, in detail
From voice notes to a second brain — capture for solo work, not meetings
/prompts/productivity — copy-paste prompts for meeting prep and post-meeting processing
/tools — note-taker tools (Granola, Otter, Fathom, Fireflies) in the directory