AI Meeting Summaries: Beyond Transcription

A transcript tells you what was said. A good summary tells you what was decided, who owes what to whom, and nothing that did not happen.

Transcript · 58 minutes
Meeting minutes
DecisionsThe launch stays on the 14th.The pilot runs in two regions, not four.
Action itemsMina sends the venue shortlist by Friday.Review the pricing page copy. Owner not stated.
Discussed, not decidedMoving the launch to the 21st was suggested.
An example of the Meeting Minutes format. When nobody was named as the owner, the summary says so instead of guessing.

Nobody reads a transcript. An hour of conversation is around nine thousand words, most of them scaffolding: greetings, restarts, somebody's dog. The thing you actually need is in there somewhere, usually in a single sentence at minute forty-one.

So the question for a meeting summary is not whether an AI can shorten text. That is easy. The question is whether you can trust what it kept, and trust that it added nothing. This post is about how we built ours around that second question.

A suggestion is not a decision

The most damaging mistake a summary can make is a quiet one: writing down something that sounded like a decision but was only an idea. Someone says "maybe we should push to the 21st", nobody objects out loud, the meeting moves on, and a careless summary records "Launch moved to the 21st."

What was said
Daniel · 00:41"Maybe we should push the launch to the 21st?"
Camila · 00:44"I'd rather not. Let's keep the 14th and cut the pilot to two regions."
Daniel · 00:45"OK. Agreed."
What the summary records
DecisionThe launch stays on the 14th. The pilot is reduced to two regions.
Discussed, not decidedMoving the launch to the 21st was suggested and not adopted.
An illustrative exchange. The idea and the agreement are recorded as two different things.

We wrote this into the summary's ground rules, in so many words. It may not invent decisions, owners, deadlines, numbers or names. It may never turn a suggestion into a decision. And it attaches an owner or a deadline to a task only when someone said it explicitly. If nobody said who will do it, the summary says the owner was not stated. An honest gap is more useful than a confident guess.

It reads what was said, not what was translated

In a translated meeting there are several versions of every sentence: the speaker's words, and a translation for each language in the room. A summary built from a translation inherits any mistake the translation made.

Ours treats the original words as the only evidence, and then writes the summary directly in the reader's language. A Korean reader and a Spanish reader each get a summary written for them from the same source, not a summary of a translation.

For names and technical terms, readers can ask to see the original spelling beside the translated word, in the sentence or as a small table. Every one of those originals is checked against the transcript, and removed if it does not appear there word for word.

Three shapes for three kinds of meeting

  • Meeting minutesTopics discussed, decisions actually made, and action items actually assigned. For any meeting that ends with work to do.
  • Lecture notesThe ideas in the order they were taught. For talks, sermons, classes and training sessions.
  • General summaryA plain account of what happened. For everything else, and for anyone who missed it.
The three summary formats.

Summaries are made when you ask for one, not automatically, and they draw on your AI allowance. Once a summary exists in a given language, the next person who opens it gets the same one instantly, at no further cost.

Long meetings are read whole

The usual way to summarise something long is to cut it into pieces, summarise each piece, and then summarise the summaries. It is cheap, and it has a specific weakness: a decision made in the middle of a long discussion depends on what came before it. Cut the meeting into pieces and the middle loses its context.

Cut into pieces, then mergedcontext lost
Read in one passthe whole meeting, start to finish
We read meetings in a single pass up to roughly 13 hours of speech. Only beyond that does the system fall back to searching and combining.

When we tested both approaches on long meetings, the piece-by-piece method lost decisions made in the middle, and the single pass kept them. So we pay for the single pass. A full-day workshop is summarised as one conversation, because that is what it was.

Ask it a question

Sometimes you do not want a summary. You want one fact. Ask AI lets you put a direct question to a meeting, while it is still going or long after, and get an answer in the language you asked in.

출시 날짜는 어떻게 결정됐나요?

출시는 14일로 유지하기로 했습니다. 21일로 미루자는 제안이 있었지만 채택되지 않았고, 대신 파일럿을 두 지역으로 줄이기로 했습니다.

  • E1 · 00:41
  • E2 · 00:44
  • 누가 장소 목록을 보내기로 했나요?
  • 파일럿 지역은 어디인가요?
A question asked in Korean about a meeting held in English. The small tags point back to the moments the answer came from.

A few details we are fond of. After a meeting, you are offered eight to ten suggested questions written for that specific meeting, in your language. In languages with polite forms, such as Korean and Japanese, the questions use them. You can ask one question across as many as ten meetings at once, which is how you find out when a decision was really made. And when an answer is drawn from a long archive, it must cite the excerpts it used.

Ask something that has nothing to do with the meeting and it will politely decline. It is there to tell you what happened in the room.

The transcript is evidence, never instructions

One last rule, which matters more than it first appears. Whatever people say in a meeting is treated as something to report, not something to obey. If a speaker says "keep this off the record", the summary does not silently drop the passage or change its own behaviour. It records that the request was made. Nothing spoken in a meeting can instruct the AI that summarises it.

Your meetings are also not used to train AI models, and the requests we send to the model provider are marked not to be stored.

Where to find it

Summaries and Ask AI are in the AI tab of the Zoom app, on each session in your web dashboard, and in the mobile app. Open a past meeting, choose a format and a language, and read the two minutes that matter out of the sixty that were said.