Meeting minutes with AI: decisions, owners and a link to the moment.

Three weeks after a steering meeting, two people disagree about what was decided. The AI summary is in everyone’s inbox. It says the group “discussed the timeline and agreed to follow up”. Nobody can tell from it what was decided, by whom, or by when.

A useful meeting record is not a prose summary. It lists decisions, owners, deadlines and open questions, each linked to the moment it was said. AI drafts it, a person approves it, and consent and retention are agreed before anyone presses record.

Why do AI meeting summaries often go unread?

Because they compress everything evenly. A prose summary gives a side remark and a decision the same weight, hedges where the model can’t tell whether something was decided (“it was suggested that…”), and rarely names an owner. Readers learn that it won’t answer their question, and stop opening it.

Summaries also fail in ways that are hard to see: an option that was discussed is recorded as decided, a joke becomes an action, or a remark is credited to the wrong speaker.

What should a useful record contain?

Four parts, each item with a timestamp:

• Decisions: what was decided, and by whom.
• Actions: each with one owner and the due date as it was stated. If someone said “by month-end”, write that; don’t invent a date.
• Open questions: what is unresolved, and who will answer it.
• Header: the meeting, the participants, the language and a link to the recording.

Picture an illustrative steering meeting for a warehouse system rollout. Before, the prose summary read: “The group discussed the rollout timeline and data migration. Concerns were raised about testing capacity. The team agreed to follow up on training.” After, the same meeting reads:

The second version takes less time to read, and every line can be checked.

Why link each item to the moment in the recording?

Because checking then takes seconds. A reviewer clicks the timestamp, hears the sentence in context and knows whether “decided” was really decided. That is the idea behind answers with receipts: people trust records they can check. The link also settles later disputes without reconstructing anyone’s memory. And it disciplines the draft: require the model to give a timestamp for every item, and treat an item it can’t tie to a moment as a likely invention, not a finding.

When Turkish and English speakers share a meeting, transcribe each in their own language, choose the language of the minutes up front, and keep the original words next to any translated decision. The link only helps if the reviewer hears what was actually said. Mixed-language audio is one place where transcription errors cluster, as the note on Turkish speech recognition explains.

A decision without a link to the moment is just someone’s memory of it.

Who approves the minutes?

For everyday meetings, the organizer reviews the draft before it is shared: a few minutes to correct owners, remove anything recorded as decided that wasn’t, and approve. Actions go into the task system only after approval.

Formal minutes, such as those of a board or committee, stay the secretary’s responsibility. AI drafts from the recording or notes; the secretary edits, approves and follows the usual process for signatures and filing. What formal minutes must contain, and whether recording such a meeting is appropriate at all, is for the secretary and counsel to decide.

Done looks like: every record shows whether it is a draft or approved, and by whom. What goes wrong: drafts forwarded as final because nobody owned the approval.

What about consent, confidentiality and retention?

Agree these before the first recording, not after a complaint:

• Consent. Participants, including external guests, know in advance that the meeting will be recorded and transcribed, and can object. Which legal basis applies is a question for your data protection officer.
• Confidential meetings. HR cases, legal discussions and deal talks are left out entirely; no recorder joins, and organizers can switch it off.
• Retention and access. Decide how long recordings, transcripts and minutes are kept, and who can hear what. When a recording is deleted, its timestamps stop working, so choose that trade-off deliberately.

Personal note-taking tools that join meetings on someone’s own account raise the same questions, so the policy should cover them too. The same questions sit behind every system we build: who can access what, and when a person decides.

How do you judge the quality of minutes?

Ask the people who were there. For a sample of meetings, ask participants: did the minutes capture every decision you remember? Is anything recorded as decided that wasn’t? Are the owners and dates right? A missed decision is the worst error, because nobody goes looking for it. Track the organizer’s edits too: if they keep correcting the same kind of item, fix the drafting, not the organizer.

Keep a small evaluation set as well: a handful of recorded meetings with minutes the organizers wrote themselves. Re-run it whenever the model, prompt or transcription changes, and count missed decisions, invented decisions, wrong owners and wrong dates.

Try it on one meeting

Pick one recurring meeting with a willing organizer, agree consent with its participants, and compare the AI draft with the organizer’s own notes for a few weeks. Keep the template to four parts. Minutes linked to the moment are one part of voice and meeting intelligence, alongside searchable recordings and call review.

This note is general information, not legal advice.