Most meetings already produce notes. The trouble is what those notes contain: a tidy paragraph per topic, a list of “next steps” nobody remembers agreeing to, and a confident line saying someone will do something by Friday. AI meeting notes can fix the first part quickly. They can also make the second part worse, because a model is very good at turning “I could maybe look at that” into an assignment with a deadline.
This guide covers a narrow job inside the much bigger topic of AI meeting notes: taking a transcript and turning it into a record your team can act on, where decisions, open questions and commitments are kept apart and every action points back to the moment it was agreed. If you are mostly shopping for a note-taking app, the capture section below gives you the short version, and our guide to choosing AI tools around the tasks you actually do is the better starting point.
Getting a transcript in the first place
You need text before you need structure. Small teams usually end up with one of four routes, and the choice matters less than people think as long as the output is a transcript with speaker names and timestamps.
Zoom’s AI Companion has a meeting summary feature that the host starts during a meeting. Worth knowing before you rely on it: Zoom says participants may automatically receive the summary after the meeting ends, which means an unreviewed draft can land in inboxes before anyone has checked it. Zoom’s getting started page says some AI features require specific licences and available AI credits, so check what your account includes.
Microsoft Teams offers live transcription that starts automatically when you record, and all participants see a notification that the meeting is being transcribed. The AI layer on top, intelligent recap, requires a Teams Premium or Microsoft Copilot licence plus transcription turned on by an admin.
Google Meet’s “Take notes for me” needs an eligible Google Workspace edition or Google AI plan, and Google recommends it for meetings between 15 minutes and 8 hours.
The fourth route is the one this article uses: download or export the transcript, then process it yourself in ChatGPT, Claude or Gemini with a prompt you control. It is slower than a built-in summary, but you decide the structure, you see the evidence, and nothing goes out until you send it.
Consent comes before the record button
Recording and transcription rules depend on where you and your participants are, what your contracts say and what your organisation allows. The tools help with notification (Zoom says it always notifies participants that a meeting is being recorded, and Teams shows a transcription notice), but a banner is not the same as agreement from a client who joined by phone halfway through. Say at the start that the call is being transcribed, explain what the transcript is for, and offer to stop. If you are pasting a transcript into a chat assistant, check whether your company policy allows client conversations in that tool at all. A legal adviser is the right person for the jurisdiction question. This article is not.
What an actionable record separates
A summary compresses. An action record classifies. The second is what you want, and it needs four buckets that most auto-generated notes blend together.
Discussion is what was talked about, including options that were considered and dropped. It is useful context and should be short. Decisions are things the group actually settled, phrased so someone who missed the meeting could act on them. Open questions are matters raised but not resolved, and they need an owner for the question even when nobody owns the answer yet. Commitments are tasks a named person explicitly accepted. Not suggested, not volunteered on someone else’s behalf, not implied by job title.
The difficult bit is the space between open questions and commitments. Spoken language is full of soft offers (“I can probably take that”, “someone should chase the supplier”) and soft dates (“early next week”, “before the install”). A good record keeps that softness visible. A bad one rounds it up into certainty because certainty looks more organised.
An illustrative transcript and where the AI got it wrong
The example below is fictional. Harbour Lane Kitchens is an illustrative six-person kitchen fitting business, and the transcript, names and output were written for this article to show a typical failure. It is not a real meeting or a test result.
| Time | Speaker | Transcript excerpt (illustrative) |
|---|---|---|
| 04:12 | Dana | The Morris job. Their worktop supplier pushed delivery again, so we need to decide whether we keep the install date. |
| 04:31 | Tom | I’d rather move it. If we fit cabinets and wait three weeks for tops, they’re living with a building site. |
| 04:50 | Dana | Agreed. We move the install. Not cancel, move. |
| 05:02 | Priya | Then the quote needs updating for the second visit. I can probably look at that, but I need the final measurements from Tom first. |
| 05:20 | Tom | I’m on site Thursday, I’ll measure then. |
| 05:34 | Dana | And someone should tell the Morrises before Friday. They booked time off. |
| 05:41 | Tom | Do we know if the supplier will cover the extra visit? |
| 05:45 | Dana | No idea. Let’s not promise anything to the client on that yet. |
Here is what a generic “summarise this meeting and list action items” prompt produced from that excerpt (also illustrative):
Action items: Priya to send the Morrises an updated quote by Friday. Tom to measure on Thursday. Dana to confirm the supplier will cover the extra visit.
Every line has a problem. Priya offered to “probably look at” the quote, conditional on measurements, and Friday was attached to telling the client, not to the quote. Nobody took the client call at all; “someone should” became Priya by proximity. And the supplier question was explicitly left open, yet the output assigned it to Dana and phrased it as if the answer would be yes. If that summary went to the client, Harbour Lane would have made a promise nobody agreed to make. The worst part is how reasonable it reads.
The corrected action log
The fix is to make the model show its evidence and to give it permission to say “not agreed”. The log below is the corrected version, with transcript references so anyone can check a line in seconds.
| Type | Item | Owner | Due | Status | Transcript ref |
|---|---|---|---|---|---|
| Decision | Morris install date moves; job is not cancelled | n/a | n/a | Agreed | 04:50 |
| Commitment | Take final measurements on site | Tom | Thursday (stated) | Accepted | 05:20 |
| Tentative | Update quote for second visit | Priya (offered, conditional) | Not agreed; after measurements | Confirm with Priya | 05:02 |
| Unassigned | Tell the Morrises about the new date | Not assigned | Before Friday | Needs owner | 05:34 |
| Open question | Will the supplier cover the extra visit? | Not assigned | n/a | Unresolved; no client promise | 05:41 to 05:45 |
This version is less satisfying to read, and that is the point. Two rows say “needs owner” or “confirm”, which tells the meeting chair exactly what to chase before circulating anything. The constraint from 05:45 (no promises to the client on cost) is preserved instead of quietly disappearing.
The prompt that produces this
The prompt below is written to stop the specific inference errors shown above. It borrows the habit from our piece on writing instructions that say exactly what output you want: define the categories, define what counts as evidence, and define what to do when evidence is missing.
You are preparing an action record from a meeting transcript.
Use only the transcript below. Do not add facts, dates or owners.
Sort content into:
1. Decisions: the group explicitly settled it.
2. Commitments: a named person explicitly accepted a task.
3. Tentative items: someone offered, or acceptance was conditional.
4. Unassigned tasks: work was mentioned but nobody accepted it.
5. Open questions: raised but not resolved.
6. Discussion: short context only, max 5 bullet points.
For every item in 1 to 5, give the transcript timestamp(s).
Owner rules: only name an owner if that person accepted the
task in their own words. Otherwise write "Not assigned".
Date rules: only give a due date if a date was stated for that
specific item. Otherwise write "Not agreed".
Keep hedges ("probably", "might", "if") visible in the item.
List any constraints on what may be told to clients or suppliers.
End with "Items to confirm before circulating".
Transcript:
[paste transcript with speaker names and timestamps]Long transcripts are where models start guessing, so for anything over an hour, process it in sections and merge the logs. If your transcript has no speaker labels, fix that first. A model asked to work out who said what from unlabelled text will produce something, and you will not enjoy auditing it.
Reviewing before the notes go anywhere
The person who chaired the meeting, or someone who attended, should review the draft. It takes a few minutes with the timestamps in front of you. Check every owner against its reference, check every date, and read the “items to confirm” list out loud to yourself. If a commitment feels surprising, it probably is.
Then resolve the gaps with people, not with the model. Message Priya and ask whether she is taking the quote. Ask Dana who calls the client. Update the log with their answers and the date they confirmed. Only then circulate. Do not let any tool send task assignments or client messages automatically from an unreviewed record; a summary that assigns work on someone’s behalf is a small act of management, and it should come from a human who is accountable for it.
If the same type of meeting happens every week, the review steps are good material for a written procedure. Our guide to creating an SOP with AI shows how to turn a repeated task like this into steps someone else can follow. And when the meeting was a messy client call where the scope itself is unclear, the Brief Decoder is a better tool for sorting out what the client actually asked for.
Meeting-note template
Copy this into your notes tool and fill it from the AI draft after review.
| Field | What goes here |
|---|---|
| Meeting | Name, date, attendees, who chaired |
| Recording and consent | Transcribed? How participants were told. Any objections. |
| Purpose | One sentence: what this meeting was meant to settle |
| Decisions | Each decision with transcript reference |
| Commitments | Task, owner (accepted in own words), due date if stated, reference |
| Tentative and unassigned | Offers, conditional tasks, “someone should” items |
| Open questions | Question, who will find out, reference |
| Constraints | What must not be promised or shared yet |
| Discussion | Up to five lines of context |
| Review | Reviewed by, date, items confirmed with whom |
| Circulated | Date sent and to whom |
Where this fits and where it does not
This workflow suits meetings where commitments matter: client calls, project check-ins, weekly operations meetings. It is overkill for a brainstorm, where a loose summary is fine and nobody expects an action log for “what if the logo was green”. It also does not replace a note-taking app if your real problem is getting transcripts at all; for that, compare the capture options above against what your existing subscriptions already include.
Direktorium covers these practical AI workflows for small businesses, one task at a time. If you would rather have someone set up the transcript, prompt and review process for your team, BBDirector can build it with you.
