Skip to content

Writing meeting minutes and notes with AI

By the KI können editorial team8 min read

Written with AI and checked against the sources by a second AI. As of November 16, 2026. More in the notes on AI.

After the meeting you have three pages of notes, and the minutes are due tomorrow. An AI brings order into them. So that in the end only what was really decided is in there, it needs clear limits, and you need a fixed routine for checking.

A team sits around a table in a meeting, a woman takes notes on a laptop, a colleague is speaking, a flip chart stands in the corner.

Key points

  • Decide beforehand which kind of minutes you need. The instruction depends on it.
  • The AI may only use what is in your notes. It marks gaps with “?”.
  • Recordings and automatic transcripts only with everyone’s consent and under your employer’s rules.
  • Where the exact wording matters, you decide, not the AI.
  • Personnel talks and health topics don’t belong in an AI chat.

First clarify: which kind of minutes?

“Write the minutes” is too vague for an AI. There are several kinds, and each has a different purpose. Put the kind into your instruction, and the format will fit.

Kinds of minutes and what the AI needs for them
KindWhat forWhat the instruction must say
Action minutesTeam meeting, regular check-inStructure: decisions, tasks, open questions
Detailed minutesCommittees, exams, contentious meetingsOrder of contributions, who said what
File notePhone call, short conversationDate, occasion, what was agreed, next step
Task listProject, construction site, office moveColumns What, Who, By when

For most meetings, action minutes are enough. They are short, and they leave less room for errors, because they only contain results.

In the trial lesson you put your first instruction together and get feedback. No account, in a few minutes. Start the trial lesson

Action minutes from your bullet points

Ruth runs a city library and writes the minutes of the monthly team meeting. Her notes are brief and in the order in which people spoke. The AI is supposed to organize them by topic and also tell her what remained unclear.

Sample instruction
Turn my notes below into action minutes of the team meeting on [date].
Organize by topic, not by the order of the notes.
Per topic: the decision in one sentence, then tasks as a table with What, Who (initials), By when.
Only use what is in the notes. Don’t invent deadlines or people responsible; write “?” if something is missing.
Separate clearly: “decided” and “only discussed”.
Below the minutes: a list of the points I should follow up on, with a short reason.
Notes:
[notes with initials instead of names]

The list at the end is worth more to Ruth than the minutes themselves. It says, for example: “Reading night: date mentioned, but nobody responsible.” Before she sends anything out, she clarifies these points with the team.

When the exact wording matters

With detailed minutes, things are different. Who raised which objection, who agreed, who didn’t? An AI likes to smooth sentences and summarize, and in the process a statement can shift. “We should check whether …” then becomes “We will …”.

Gregor manages IT projects at an energy utility. In the steering committee of his current project, it is disputed whether a deadline will be postponed. He has the AI handle only the format and says explicitly that it must not rephrase statements.

Sample instruction
Put my notes below into the format of detailed minutes.
Per agenda item: contributions in order, who (initials) said what, result.
Don’t rephrase the contributions. Only correct spelling and write full sentences where I have keywords.
Where my notes are unclear, mark the spot with [unclear] instead of interpreting it.
Notes:
[notes with initials]

Then Gregor reads every contribution against his notes. Before the minutes go out, he sends disputed passages to the people involved and asks them to confirm.

Recording and transcripts: what is allowed

Many meeting programs can transcribe and produce a summary from it. That saves work, but it raises questions. Under Section 201 of the German Criminal Code (Strafgesetzbuch, StGB), anyone who records a non-public conversation without authorization commits a criminal offense. On top of that come data protection and your employer’s rules.

  1. Clarify whether your employer allows recording, transcription or AI summaries for meetings, and with which program.
  2. Ask all participants beforehand, including outside guests. A silent icon in the corner isn’t enough.
  3. If anyone says no, the recording stays off.
  4. Note who agreed, for example in the first line of the minutes.

Some programs offer a middle way. In Microsoft Teams, the organizer can set Copilot to run only during the meeting, without a saved transcript. Anyone with a license can then ask questions and have notes made; afterwards, no summary appears in the meeting recap. Copilot still processes what is said, though, and the company can keep questions and answers according to its own rules. So let the participants know in this case too. Whether the setting is enabled at your workplace is decided by IT.

If there is a works council (Betriebsrat), the elected employee representation, it usually has a say in such functions, because they can show who said what and when.

Turning a transcript into minutes

Nathan is a site manager at a building construction company. His employer allows transcription in the weekly site meeting, and everyone involved has agreed. The transcript is long, contains filler words and sometimes wrong names, because the speech recognition mixes up voices. He uses the AI his employer has approved.

Sample instruction
Here is the transcript of our weekly site meeting on [date].
Create action minutes: decisions, tasks (What, Who, By when), open points.
For every decision and every task, give the timestamp in the transcript so that I can listen to it again.
If it isn’t clear who said something, write “speaker unclear”.
Leave out small talk and repetitions.

The timestamps are his checking tool. For every decision, he jumps to the spot and reads or listens again. Speech recognition errors, for example in figures and measurements, stand out this way before they end up in the minutes.

A file note after a phone call

Not every conversation is a meeting. Joel is a dispatcher at a trucking company and talks to customers on the phone every day about delays and loading times. He records what was agreed in a short file note so that his colleague on the late shift is in the loop. He types his bullet points right after the call.

Sample instruction
Turn my bullet points into a file note on a phone call, at most 80 words.
Structure: date and time, the other party as [customer A], occasion, what was agreed, next step with who is responsible.
Copy times, quantities and loading points exactly as they appear in the bullet points.
Bullet points:
[bullet points]

Joel only adds the real customer name in the dispatch software. And he reads the note once against his bullet points, because a mixed-up time here sends a truck out for nothing.

Tasks, deadlines and who is responsible

The part of the minutes that gets read the most is the task list. This is where an error does the most harm: a wrong deadline or a task that belongs to nobody.

  • Every task has exactly one responsible person, not a group.
  • Deadlines as a date, not as “soon” or “next week”.
  • If a detail is missing, it says “?” and you ask instead of guessing.
  • Tasks from the last minutes that are still open go back on the list.

The last rule is one an AI can handle well. Give it the task list from the last minutes and the new notes, and ask it to carry over open points and mark them as “from the meeting on [date]”.

What doesn’t belong in AI minutes

Carmen leads a team in the customer service department of an insurance company. She has the company AI organize her team meetings. She doesn’t put notes from one-on-one talks with employees into it, not even with initials.

  • Personnel talks, formal warnings, performance reviews
  • Illnesses and other health information
  • Conflicts between people
  • Internal meetings of the works council
  • Anything your employer classifies as confidential, such as prices or acquisition plans

You write notes like these yourself. More on what may go into which AI is in the guide Data protection and AI tools.

The last round before sending

  1. Compare every decision with your notes or the timestamp in the transcript.
  2. Clear up every “?” and [unclear] or leave it marked as open, never guess yourself.
  3. Check names and initials; speech recognition likes to mix them up.
  4. Look up figures, measurements, amounts and dates one by one.
  5. Send the minutes with a deadline for corrections, such as “Objections please by [date]”.

How to practice this

Whether it’s a team meeting or a site meeting: the instruction tells the AI what it may organize and what it may not. In the learning app “KI können”, you phrase limits like “invent nothing, mark gaps” with examples from many professions and try them out right away in your own AI. To get a taste, there is a trial lesson that doesn’t require signing up.

Sources

Information as of November 16, 2026. Laws and offers can change; this article is not legal advice.