An AI note taker works in two steps: speech recognition turns the meeting into a transcript, and a language model turns that into a summary with decisions and action points. You choose between your meeting platform, a standalone meeting assistant or a recording you have transcribed and written up yourself.
The quality of the minutes depends on the prompt and on the mistakes the model can make. A proposal easily becomes a decision, and an action point without an owner sometimes gets a made-up deadline.
Checking is therefore not a formality, and neither are the rules around recording: criminal law, the GDPR, the works council and the ban on emotion recognition in the workplace.
For some of your meetings, AI is not the right choice, for example in confidential conversations about individuals.
How does an AI note taker work?
Taking minutes with AI works in two steps: speech recognition turns the audio into a verbatim transcript, and a language model turns that transcript into a summary with decisions, agreements and action points. The second step is never better than the first.
The summary is never better than the transcript it is based on.
Poor audio, participants talking over each other, dialect and jargon produce errors in the transcript, and the language model builds on them. If you already use AI summaries as a way of working smarter with AI, you will notice that difference first in meetings with many speakers.
What is the difference between a transcript and minutes?
A transcript is a verbatim record of what was said; minutes are an approved record of decisions and agreements. An AI tool delivers the first and a summary, but not minutes in the formal sense.
The table below compares the three forms, so you can see what an AI tool delivers and what a person still has to do.
| Form | What it is | Who makes it | Status |
|---|---|---|---|
| Transcript | Verbatim text of the recording, per speaker | Speech recognition | Raw source |
| Summary | Key points, decisions and action points | Language model | Draft |
| Minutes | Record of what was discussed and decided | Minute taker or chair, possibly based on the summary | Approved |
At a board meeting, a works council meeting or a general meeting of members, the next meeting approves the minutes. In those situations, an AI summary is always a draft that someone with responsibility still has to approve.
Which AI note taker options are there?
There are three AI note taker options: the feature built into your meeting platform, a standalone meeting assistant and a recording you have written up yourself. The routes differ mainly in where your meeting data ends up.
How do you take minutes with AI in Teams or Zoom?
Microsoft Teams, Zoom and Google Meet transcribe online meetings and produce an AI summary within their own environment. The data therefore stays in an environment covered by agreements your organisation already has. Which summary features are available depends on your licence; for Teams, this runs through Microsoft Copilot. The limitation: this route only works for meetings held on the platform.
What does a standalone AI note taker do?
A standalone AI note taker, such as Otter or Fireflies, joins the call as a bot or browser extension and works across several platforms. The downside is an extra party processing your meeting data, so you need a data processing agreement and must know where the recordings are stored. A note-taking bot that an employee installs without an agreement with the organisation is a textbook example of shadow AI.
How do you take minutes of an in-person meeting with AI?
You take minutes of an in-person meeting with AI by recording it on a phone or recorder, after which a transcription model converts the audio to text. OpenAI’s Whisper is one such model and can run locally, so the recording never leaves your own computer. You then give the transcript to a chatbot with a fixed prompt. This route gives you the most control over your data, but it also takes the most manual work.
Which prompt do you use for minutes from a transcript?
A good prompt for minutes from a transcript keeps decisions, discussion points and action points apart and has the model indicate what it is unsure about. A model that only gets “make a summary” mixes those three and fills in gaps itself.
The following prompt enforces that separation:
Below is the transcript of a meeting, with the agenda and the list of participants. Write a report in three parts. Part 1: decisions, covering only what was explicitly decided, each with the quote from the transcript you base it on. Part 2: points discussed without a decision. Part 3: action points with owner and deadline. If an owner or deadline is not in the transcript, write “not agreed”. Do not make up names, amounts or dates.
Quoting each decision makes checking quick afterwards, because you only have to reread those passages. The agenda and list of participants are the context the model needs to spell names correctly and sort points properly; how you give context to an LLM determines much of the result.
What mistakes does AI make when taking minutes?
AI makes five recurring mistakes when taking minutes:
- Wrong speaker: a statement is attributed to the wrong participant, especially when people talk over each other or are off camera.
- Proposal as decision: “shall we increase the budget?” ends up in the report as a decision, even though nobody agreed.
- Missed negation: a “no, unless” or “not before the third quarter” disappears, so the agreement says the opposite.
- Made-up action point: the model fills in a missing owner or deadline itself, because an action point without those fields looks incomplete.
- Garbled names and jargon: client names, product names and abbreviations are spelled as they sound, and wrongly, yet still look plausible.
The made-up deadline is a form of AI hallucination: the model generates the most probable continuation, not the correct one. In the period when I worked intensively with ChatGPT, I consistently came across summaries that sounded plausible but differed in substance from the source.
How do you check minutes that AI has produced?
You check minutes that AI has produced by checking every decision and every action point against the transcript before the report is circulated. Reading the summary without the transcript next to it is not a check.
- Compare each decision with the quote or passage in the transcript it is based on.
- Check names, amounts and dates, because they still sound convincing after a speech recognition error.
- Remove action points for which there is no owner or agreement in the transcript.
- Have the chair or a second participant approve the draft before it gets the status of minutes.
From my years at major banks, I bring the four-eyes principle to my AI work: nothing goes out of the door without a second check. For me, a final human edit is not an optional step but part of the way I work, and that applies all the more to minutes that record agreements.
Can you record a meeting for AI minutes?
Yes, you can record a meeting for AI minutes, as long as you record as a participant, inform the other participants in advance and process the recording in line with the GDPR. Three frameworks are involved; the examples below are from Dutch law.
Under criminal law in the Netherlands, recording is permitted under Article 139a of the Dutch Criminal Code if you take part in the conversation yourself. If you have a note-taking bot attend a meeting you are not part of, that permission does not apply.
The GDPR asks more of you. A recording contains personal data, so you record the purpose, announce at the start that the meeting is being recorded, set a retention period and conclude a data processing agreement with the tool’s supplier. If an organisation systematically records meetings with employees, a DPIA is the obvious step.
For employers, the works council comes into it too. In the Netherlands, a facility suitable for monitoring employees’ attendance, behaviour or performance falls under the right of consent in Article 27 of the Dutch Works Councils Act (Wet op de ondernemingsraden), and a meeting tool with speaking-time or participation statistics can fall under it.
Which AI features in meeting tools are prohibited?
One feature of meeting assistants that has been prohibited since 2 February 2025 is emotion recognition in the workplace. The AI Act prohibits systems that infer employees’ emotions from biometric data such as voice or facial expression, with an exception for medical and safety purposes.
The distinction lies in the source. A mood or engagement score derived from tone of voice or video falls under the ban; an analysis of the tone of written text does not. If a meeting assistant offers an engagement or mood score per participant, switch it off until it is clear what it is based on.
When is it better to take minutes without AI?
It is better to take minutes without AI in conversations involving sensitive personal data or a duty of confidentiality. Examples are performance reviews and absence meetings, discussions of medical or legal issues and reorganisation meetings before anything has been announced.
In those conversations, the risk of a recording held by an external party outweighs the time saved. The same goes for a meeting where nobody takes the time to check the transcript: an unchecked AI report is then worse than brief handwritten notes.
Frequently asked questions about AI note takers
With an AI note taker, much depends on the tool your organisation already has, and that is what the questions below are about.
Can Teams transcribe a meeting in languages other than English?
Yes, Teams can transcribe meetings in languages other than English; Dutch (Netherlands) and Dutch (Belgium), for example, are on the list of supported spoken languages. You choose the language in the transcription settings, under Spoken language in this meeting. Set it to the language the participants speak, so that speech recognition works in the right language.
Can you use ChatGPT to take minutes of a meeting?
Yes, ChatGPT can take minutes of a meeting with record mode, which transcribes and summarises recordings of meetings or voice notes. Since 29 September 2026 there has also been the Meetings plugin that takes notes for you, and since 6 October 2026 you can upload audio files to have a transcript and summary made of them. Record mode only works in the desktop app for macOS and is available on Plus, Pro and the business and education plans. According to OpenAI, the audio is deleted after transcription and record mode works best in English. Asking the other participants for consent is your responsibility.
Can you use Copilot in Teams without recording the meeting?
Yes, you can use Copilot in Teams without recording the meeting if the organiser chooses Only during the meeting under Copilot and other AI in the meeting options. No transcription then starts, and Copilot is not available in the Recap tab afterwards. So copy anything you want to keep while the meeting is still running. Your prompts and Copilot’s answers may still be retained under your organisation’s retention policy.
How do you learn to use AI for minutes well?
You learn to use AI for minutes fastest by first understanding what a language model does with a transcript, and then using one fixed way of working: the same prompt, the same check and the same agreements on recording. A LearnLLM AI course such as the AI Literacy Course covers that foundation, including writing prompts and checking output, before you choose a specific tool. That way, AI gets your minutes done faster without meetings being captured in a report nobody has checked.
