With the right licence, Teams, Zoom and Meet create meeting notes automatically. Depending on the provider, you enable transcription, Meeting Summary or "Take notes for me", and the feature may first need admin approval. Microsoft Teams produces AI notes and recommended tasks. Zoom's AI Companion produces summaries, key takeaways and action items with speaker attribution. Google Meet writes quick notes or full notes with decisions, action items and details. Across platforms, a standalone service such as Otter, Fathom, tl;dv or Notion fits better. But the output only becomes usable through structure: decisions, tasks with an owner and a deadline, open points with a date. And one thing comes before all of it: recording in secret is a criminal offence in Germany, and the safe house rule is prior agreement from everyone involved.
I stopped writing minutes by hand a long time ago, but I do not let them run unattended either. The automation handles writing while listening, the most expensive and most error-prone part. What counts as a resolution and what was merely an idea, I decide. How I take a workflow apart before I automate it is on Process optimisation, and which tasks are suited to being handed over is covered in Which tasks you can hand to AI.
A summary and minutes are not the same thing
Almost every tool advertises a "summary". That is honest, but not what you need after a meeting. The difference only shows up three weeks later, when someone asks whether that was actually decided.
The vendors describe it this way themselves. Zoom lists its output as "AI-generated summaries, key takeaways, and action items". Microsoft calls the intelligent recap output "AI meeting notes" and "AI recommended tasks". Only Google Meet lists "decisions" as a section of its own in full notes. No vendor promises you a record of resolutions. If you want one, you have to supply the structure, for instance with the instruction from Have Minutes Written: ChatGPT, Teams, which turns a transcript into a record of decisions.
The structure that holds up
Minutes that people actually use later have five blocks, in an order that deliberately does not follow the conversation.
- Header. Date, occasion, participants, duration. So it is clear later who was there.
- Decisions. One present-tense sentence each, describing what applies from now on. If you catch yourself writing "we should", it is not a decision but block four.
- Tasks. A table with three columns: task, owner, deadline. Without a name it is not a task but a wish. Without a date it is not a deadline but an intention.
- Open points. Everything that was not decided, plus when it will be decided and by whom. This block is the most valuable one: it answers the first question of the next meeting.
- Context. Only what explains a decision. Everything else stays in the transcript.
What does not belong: the flow of the conversation. A running record is a transcript with paragraphs, and you already have that. Append your tool's own summary text at the end rather than putting it on top.
The fastest route is to give the structure to the tool. Google Meet lets you customise the sections of full notes, and Otter, Fathom and Notion work with their own summary templates. If you drop transcript and raw summary into an AI instead, write the instruction once and save it, see Claude prompts that work in daily use.
What the built-in tools do, as of September 2026
The three big conferencing tools cover the standard case, as long as everyone sits on the same platform; how to switch on transcription in Teams step by step is in Transcribing Calls in Microsoft Teams.
Two details from Microsoft's documentation bite in daily use. For Teams events, only organisers, co-organisers and presenters see the recap; attendees do not. And video recap requires a recording and a meeting of at least ten minutes; transcription alone is not enough.
Standalone services, when the platform keeps changing
If you talk sometimes in Zoom, sometimes in Meet and sometimes on the phone, built-in tools turn into patchwork. That is when a service that hooks into all three pays off.
Prices in both tables are vendor figures in the currency and billing mode named, as of September 2026, retrieved on 19 September 2026; monthly prices and annual prices converted to a month are listed separately. Some include promotional discounts, and vendor prices change; check the vendor's page before you buy.
More important than the price is the question: where does the recording end up? A service that transcribes on your instructions is a processor, and if it processes outside the EU, third-country transfer rules apply on top. What data goes towards AI in my setup is in AI and data protection: what the AI is allowed to see.
One paragraph on consent, before you record
Automatic minutes require a recording or a live transcription, and recording the non-publicly spoken word without authorisation is a criminal offence under § 201 of the German Criminal Code, not an administrative fine. What authorises it is prior agreement from everyone else whose voice lands on the recording, not just your main counterpart. For everyday business that is the safe house rule; whether there are narrowly limited exceptions beyond it is a question for the individual case and belongs with a lawyer. The practical route is a sentence at the start of the call naming purpose and retention period and allowing a consequence-free no, plus a note of who agreed. The reasoning with sources, including the GDPR side and the special case of employee conversations, is in Transcribing calls: what German law says. I am a developer, not a lawyer; this is a working basis, not legal advice.
The process in six steps
- Clarify and document consent. Before the first recording, not before the tenth.
- Switch on transcription, not just recording. In Teams the entire recap depends on transcription being allowed by policy and active in the meeting.
- Store your minutes structure as a template. Write it once, then use the same shape for every meeting.
- Book follow-up time right after the meeting. Separate decisions from ideas, put names on tasks, set deadlines. I plan ten minutes for it, as my own rule of thumb for planning, not a measured figure.
- Distribute while it is fresh. Minutes that sit for three days do not get corrected, they get ignored.
- Set a retention period and delete. The raw recording and the finished minutes get separate deadlines. Teams keeps audio recaps for only 60 days in the user's OneDrive, but that is no substitute for your own deletion policy.
Where the automation reliably falls short
Six places where you will follow up, whichever tool you pick.
- Decision or idea. Whether "we should look into that" was an assignment depends on the context of the conversation, not the wording. That is why Microsoft deliberately calls the output "recommended tasks".
- Assigning an owner. A name is only picked up if it was actually said. "Can you take that?" produces no owner in the minutes.
- Proper nouns and jargon. That vendors sell vocabulary lists as a paid feature, Otter from Pro for instance, shows where recognition struggles: product and company names and jargon.
- Switching languages mid-call. Google states outright that multiple languages in the same meeting are not supported, and in Teams the multilingual recap is currently in public preview. A call that jumps between German and English is the standard case in many teams and the edge case in the documentation.
- Too short and too long. Google recommends 15 minutes up to a maximum of 8 hours and names incomplete summaries as the risk for shorter meetings; Microsoft skips transcripts that are too short in audio recaps.
- Errors stay errors. Microsoft states explicitly that audio recaps are AI-generated and may contain inaccuracies. Distribute unread and you distribute them with your name underneath.
That leads to a useful rule: automatic minutes save you the writing, not the thinking. Plan the follow-up in instead of hoping you will not need it.
From single minutes to a pattern register
This is where a minutes tool reaches its limit, and where the part that pays off for freelancers and consultants begins. Minutes answer the question: what did we discuss? After fifty client calls you have fifty answers to that and none to the far more valuable question: which question do my clients keep asking, and what is my best answer to it?
That second question is exactly what Gustav, my AI employee for evaluating call transcripts, works on. He compares many conversations instead of summarising one, and condenses them into a pattern register: the recurring questions, each paired with the best answer you have already given yourself. Out of that come a sales playbook, the FAQ of your website, raw material for posts, and an objection library. In my setup, Gustav uses Whisper locally on the machine for speech recognition. The anonymisation pass described here comes afterwards. He distils what you said yourself and does not invent advice on top.
The side effect is convenient in data protection terms: what remains is not an archive of personal recordings but a register of roles, industries and ranges. As a minimum test, the anonymisation pass checks whether role, industry and a rare piece of content together still point back to a person. True anonymity depends on the full context and all means reasonably likely to lead back to that person. If doubt remains, the entry is pseudonymised and gets a purpose, a deadline and protection like any personal filing. How such a role comes into being is described in Hiring an AI employee: the process, I collect the others under AI employees, and which carry weight in coaching work is in AI for coaches.
Frequently asked questions
Can AI create meeting minutes automatically?
Yes, with the right licence: Teams, Zoom and Google Meet generate notes, summaries and task suggestions. Depending on the provider, you enable transcription, Meeting Summary or "Take notes for me". That is a summary with task suggestions, not a record of resolutions: separating decisions from ideas and assigning tasks to people remains manual work.
What do automatic meeting minutes cost?
Zoom includes AI Companion in paid Workplace services at no extra cost, Microsoft adds Copilot from $18.00 per user per month billed yearly, and with Google the entry point for the note-taking feature is Business Standard at €13.60 per user per month, because Gemini in Meet is only included from that edition. Fathom and Otter have free tiers. Vendor figures, as of September 2026.
What separates AI meeting notes from a real evaluation?
Notes summarise one conversation. An evaluation compares many and shows which question recurs how often and which answer works best for you. The first is handled by a conferencing tool, the second by a fixed process across the whole body of calls, with anonymisation in front of it.
Can I simply have a meeting recorded?
Not in secret. Recording the non-publicly spoken word without authorisation is a criminal offence under § 201 StGB. The safe house rule is therefore: ask beforehand, and ask everyone involved. The details, including the GDPR side and the exceptions, are in the post on the legal situation.
How to carry on
Take your next meeting and do two things: switch on transcription after asking, then spend ten minutes writing the three blocks yourself, decisions, tasks with names and deadlines, open points. The gap between your version and the tool's is precisely the template you give it from then on.
If you then want to know which patterns sit in your conversations, Gustav is available as a finished package in my Community, with the anonymisation pass, consent templates and an invented demo transcript to practise on. And you do not build your own minutes template alone there: others share their version, and in the calls we go through yours together. The templates are samples without warranty and do not replace a lawyer's review of your case.
