Meeting minutes with AI: from recording to sign-off

With AI, a recording becomes a draft of the minutes, not finished minutes. You save time mainly in preparation and in writing up. The meeting lasts as long as before, and sign-off stays with you. Sources as of November 2026.

What a set of minutes costs today

On the author’s working assumption, a set of minutes takes about 240 minutes: 30 minutes of preparation, 90 minutes of note-taking during the meeting and 120 minutes of writing up. This is a guide figure for project leads, not a measurement. I am not aware of any independent measurement of this value.

With AI only two items change. By my assumption, preparation falls by 30 to 50 per cent, because open points and the agenda come from the last minutes. Writing up falls by 60 to 75 per cent, because the draft comes from the recording. If I calculate with the midpoints, that is 40 and 67.5 per cent, and 18 minutes of preparation and 39 minutes of writing up remain. The 39 minutes include checking, correcting and signing off.

Together that is 147 instead of 240 minutes, a calculated saving of 93 minutes or 39 per cent. The range runs from 81 to 105 minutes. I do not count note-taking as a saving, because it is time spent in the meeting.

The values are assumptions and have not been measured. The recording brings a further benefit in the meeting itself: if you do not have to take notes, your head is free for chairing. The 90 minutes of meeting remain, but they then belong to decisions, dates and open points.

todaywith AI

• PreparationAssumption: 40 per cent less

  today 30 min
  with AI 18 min

• Note-takingstays meeting time

  today 90 min
  with AI 90 min

• Writing upAssumption: 67.5 per cent less, incl. review

  today 120 min
  with AI 39 min

Source: Author's working assumption, calculated with the midpoints of the ranges

Text version of the graphic

• Preparation: today 30 min, with AI 18 min (Assumption: 40 per cent less)
• Note-taking: today 90 min, with AI 90 min (stays meeting time)
• Writing up: today 120 min, with AI 39 min (Assumption: 67.5 per cent less, incl. review)
• Source: Author's working assumption, calculated with the midpoints of the ranges

The process in four steps

1. Recording: only after notice to everyone involved and with their consent. Otherwise you write the minutes as before. For speaker separation, the service I read (Microsoft, Fast Transcription) requires a single-channel recording.
2. Transcript: speech recognition turns the audio into text. Speakers appear as numbers, and you assign the names yourself. Typical sources of error are numbers, proper names, technical terms and noise, for example in a site cabin.
3. Draft: a language model fills your minutes template with topics, decisions, actions (owner and date) and open points. Each entry gets a time stamp pointing into the transcript. The instruction is to take over only content from the transcript and to mark anything unclear as open.
4. Review and sign-off: you compare figures, deadlines and names with the transcript at the time stamp, correct what is wrong and send. After that you delete the recording and transcript once the period you set beforehand has passed. The minutes stay.

1. Recording You, with the consent of everyone involved
2. Transcript Speech recognition, speakers as numbers AI prepares
3. Draft Language model fills your template, with time stamps AI prepares
4. Review and sign-off You, against the transcript at the time stamp

Text version of the graphic

• 1. Recording: You, with the consent of everyone involved
• 2. Transcript: Speech recognition, speakers as numbers (the AI prepares)
• 3. Draft: Language model fills your template, with time stamps (the AI prepares)
• 4. Review and sign-off: You, against the transcript at the time stamp

How to get a good draft

An entry in the draft might look like this (example with invented data): “Action: submit the structural stability calculation for the slab above the second floor. Responsible: Example Structural Engineers. Date: 19 February. Time stamp 00:41:12.” You listen to the passage at the time stamp and see at once whether the owner and the date were really named like that.

In my estimate, the quality of the draft depends more on the recording and the template than on the model. A short round of introductions with names at the start helps you assign the speakers later. A microphone in the middle of the table does more than a phone in a jacket pocket. Many services also let you supply a list of project and technical terms that they prefer to recognise. Keep that list up to date, otherwise the transcript gets the same thing wrong in every meeting.

Preparation can be supported too: from the last minutes the AI pulls the open points, with owner and date, into the new agenda. What matters today and in what order it comes up, you add and weight yourself.

A fixed order helps with checking. Decisions first: was the decision made as written? Then actions: are owner and date right? Last come figures, dimensions and names. Jump to the time stamp and listen to the passage instead of reading the whole transcript.

Whether a solution of your own is worth it depends on how many sets of minutes you write. If you write only a few a month, in my estimate a ready-made tool is enough. That is an assumption, not a measurement. Also check whether features of your video conferencing software are already sufficient.

What the AI cannot do

Whether “We’ll have a look at that” was a commitment, a statement of intent or politeness is for the person who was in the room to decide. The same goes for whether a sentence belongs in the minutes at all. Irony, glances, conversations outside the door and what was shown on the plan are not in the audio.

The draft also does not know about agreements made before or after the recording, for instance on the way to the car or in a phone call the next morning. If they apply to the project, they belong in the minutes, and only you know about them. Plan the review for the same day or the next, while the meeting is still fresh.

The model can name an owner only if someone mentioned one. If the name is missing, the entry should stay open. A model that fills gaps itself produces actions that nobody agreed.

The minutes remain your document

Minutes record what was agreed. In a dispute about deadlines, commitments or notices of defects, they can play a role. What legal effect they have in an individual case is for your lawyer to clarify. The Federal Chamber of Engineers (Bundesingenieurkammer) states that when AI tools are used, in the end it is always engineers who take personal responsibility. The statement concerns AI-based services in engineering in general; for minutes it applies by analogy. Whoever signs off the draft has read it.

A practical sentence at the end of the minutes helps: objections by a stated date. It brings errors to light sooner than the next meeting. What effect a missing objection has in a dispute is for your lawyer to clarify.

When you choose platforms for processing, look for operation in the EU or locally and for a data processing agreement (Auftragsverarbeitungsvertrag). Such a workflow can be operated in line with the GDPR, but that is not a guarantee. Your data protection officer (Datenschutzbeauftragter) clarifies the individual case, and your lawyer clarifies the recording itself.

What you can do now

1. Time preparation, note-taking and writing up separately for three sets of minutes. Only then will you know whether the assumptions in this article fit your practice.
2. Write down your minutes template: which fields belong in it, and how are actions and dates tracked?
3. Before the first recording, settle the rules for consent, deletion period and storage location.
4. Try it first in an internal meeting, with everyone’s agreement, and compare the draft with your own notes.