AI Meeting Notes: Tools, a Prompt That Works, Consent

Writing up a meeting takes longer than it should, and the writing is not the slow part. Remembering is. Who took the action, by when, and was that a decision or just someone thinking out loud? That’s precisely the part an AI handles well: it listens to everything and sorts it afterwards.

The output, though, depends on two things that happen before the meeting starts. Everyone has to know the call is being transcribed, and you need a clear template for what the notes should contain.

How AI meeting notes are produced

There are always two separate steps. Speech recognition turns audio into text. Then a language model turns that text into notes. Some tools hide this behind one button, but it pays to keep the steps apart in your head, because errors creep in at different points.

1. AnnounceSay at the start that the call is transcribed; let people object.
2. TranscribeInside the meeting tool, or from a recording converted afterwards.
3. StructureSummarise the transcript with a fixed template.
4. CheckCompare decisions, names and deadlines with the transcript.
5. Share, then deleteSend the notes; delete raw audio after an agreed period.

Speech recognition stumbles on jargon, unusual names and people talking over each other. The language model makes a different mistake. It likes to turn “maybe we should look at…” into a firm decision. That’s why the check step belongs in the process, not as an afterthought.

Which tools do it

If your team already lives in Teams, Google Meet or Zoom, the built-in option is usually the path of least resistance.

Microsoft lists summarising and transcribing meetings plus capturing action items among Copilot’s Teams features, with the depth depending on licence. The tiers are laid out in Microsoft’s Copilot overview.

Google Meet has “Take notes for me”. Per Google’s help page, it needs an eligible Workspace edition or Google AI plan, saves notes to Google Docs, and supports eight languages including English, French, German and Spanish. One language per meeting, though.

Zoom’s product page lists AI meeting summaries even on the free Basic tier, capped at three hosted meetings a month. Anything beyond that means a paid plan.

For in-person meetings without a video platform, there’s the phone on the table. You convert the recording afterwards, either with a cloud service or locally with an open-source speech model such as Whisper. Running it locally keeps the audio on your machine. It is slower and needs some technical comfort.

The prompt that produces usable notes

Paste a transcript with “summarise this” and you get a school-essay paragraph. Notes need structure. This template works in any chat assistant:

Below is the transcript of our team meeting on 14 October.
Turn it into meeting notes with exactly these sections:

1. Decisions: only what was explicitly agreed, one line each.
2. Action items: table with Task | Owner | Due date.
   If no date was given, write "not set".
3. Open questions: raised but not resolved.
4. Next meeting, if mentioned.

Rules:
- Never invent owners or due dates.
- If it is unclear whether something was decided, list it under
  "Open questions" and tag it [check].
- No narrative of the discussion. Outcomes only.

<transcript>
...
</transcript>

Two lines do most of the work. “Never invent due dates” stops “soon” from becoming a specific Thursday. The [check] tag sends your eyes straight to the places where the model was unsure. Anthropic’s prompting guide also recommends placing long documents like transcripts at the top of the prompt, above the instructions. You’ll find more of these techniques in our guide on how to write prompts that work.

What a good result looks like

Here’s a made-up but typical example. Four people discuss relaunching a company website. After 40 minutes, the template above returns roughly this.

Below the table sit two decisions (“launch moves to January”, “the old blog won’t be migrated”) and one open question about budget. Nothing else. That’s the point. Notes that fit on one screen get read; three pages of who-said-what don’t.

The [check] tag isn’t a bug. It’s the behaviour you asked for. In the transcript, Lea said she’d “have a look at it”. Whether that was a commitment is a human call, ideally settled with a quick message to Lea before the notes go out.

A second pass often pays off. After the first draft, ask: “List every place in the transcript where a deadline or a person’s name is mentioned, with timestamps.” Now you can verify the action list against the source without rereading the whole transcript.

Consent and the law

This part is not optional. Recording rules vary by country and, in the US, by state. California, for instance, requires the consent of everyone on the call, while many other states only need one party to agree. Germany treats secretly recording a private conversation as a criminal offence. For a team that spans borders, the safe default is simple: tell everyone, every time.

Across the EU and UK, the GDPR applies on top, because voices and what people say are personal data. You need a purpose, transparency, a retention period, and a data processing agreement with any cloud vendor. German workplaces with a works council also need its agreement before introducing tools that can monitor performance, and transcription tools typically qualify.

• Everyone told at the start, with a chance to object
• External guests asked explicitly
• Data processing agreement in place with the vendor
• Works council or employee reps consulted where they exist
• Retention period set for audio and raw transcripts
• No transcription of HR or health conversations

We’d keep two kinds of meeting off limits entirely: HR conversations and anything touching health. The benefit is small, and GDPR Article 9 puts health data in a special category. Our piece on ChatGPT and the GDPR covers the trade-offs in more depth.

Where AI notes go wrong

Treat the output as a draft, not a record. These errors come up again and again:

• Names misspelled or misattributed, especially on calls from a shared room microphone.
• Sarcasm and hedging lost. “Sure, we’ll have it by tomorrow” with an audible sigh becomes a commitment.
• Numbers swapped, such as 15 for 50.
• Gaps in the transcript smoothed over instead of flagged.

That last one is the sneakiest. Fluent notes feel complete. So the person who ran the meeting should read them before they go out. For an hour-long meeting that’s a few minutes of reading instead of half an hour of writing; treat that as a rough guide, not a measured figure, since it depends heavily on how your meetings run.

When the notes should feed your systems

It gets interesting when notes don’t just land in an inbox. Creating tasks in your project tool, filing decisions in the wiki, putting deadlines into calendars: standard tools only manage part of that. Teams that want it wired into their own stack can build it with a development partner such as OnlineBuilders.

To start, you need less. Take the template above, announce at your next team meeting that you’re trialling transcription, and compare the AI notes with what people remember. If deadlines from those notes should go straight into calendars, our guide to AI scheduling assistants picks up from there.

Questions people ask

Do I need consent if the tool only takes notes and doesn't record?

Live transcription still processes what people say, and under GDPR that's personal data. Announce the tool at the start and give people a chance to object. Many tools also show a visible banner for this reason.

Can AI tell who said what?

Teams, Meet and Zoom attribute speech to signed-in participants. A phone recording from the middle of a meeting room is much worse at separating speakers. Saying names out loud helps.

How long should I keep transcripts?

Only as long as you need them. For most teams that means saving the approved notes and deleting the audio and raw transcript once the notes are signed off.

Does it work for meetings in two languages?

Poorly. Google states that Meet's note-taking supports one language per meeting. Mixed-language calls produce patchy transcripts.