Everyone in the meeting relaxed the moment the recorder joined. Nobody wrote anything down, because it was being captured.
Two weeks later you need to know what was agreed about the deadline. There is a 6,000-word transcript and a summary with six bullets, none of which is the thing you need, and searching the transcript returns eleven mentions of "deadline" in fragments of sentences with no way to tell which one was the decision.
The technology worked exactly as advertised. Capture is solved. It turns out capture was not the problem.
What notes were for
Taking notes was never primarily about recording what was said. It did three things, and only the first is what a recorder replaces.
It kept a record. Genuinely replaced, and improved on — a transcript is more complete and more accurate than anything a human types.
It forced attention. You cannot take notes without deciding, continuously, what matters. That decision is the work, and it is why people who take notes remember meetings better even when they never reread them. A recorder removes the decision, which is the point and also the cost.
It produced an artefact you could use. Six lines you can read in twenty seconds before the next meeting. A transcript is not that. It is a source document, and reading it costs about as long as the meeting did.
The gap is between record and artefact, and no amount of transcription quality closes it. A better transcript is a better source document.
What the summaries do and do not do
Automatic summaries are genuinely useful, and it is worth being precise about the boundary.
Good at: who was there, what topics came up, the rough shape of the discussion, and a list of anything explicitly stated as an action.
Unreliable at: what was actually decided as opposed to discussed, what was agreed implicitly, who is really doing something when nobody said "I will", the difference between a firm commitment and someone thinking aloud, and — most importantly — the thing that mattered to you, which was probably one sentence somebody said in passing.
That last one is the fundamental issue. A summary is a general-purpose compression of the meeting. What you needed was a specific extraction relative to your own concerns, and the model does not know what those are.
The three lines that make the recording useful
Keep the recording. It is a better record than you would produce. Then, within ten minutes of the meeting ending, write three lines.
What was decided. In your words, plainly. "We are moving the launch to the 12th."
What I owe, and to whom, by when. Yours specifically. It becomes a dated task straight away — one sentence with the date in it — because an owed thing that lives only in a summary is not tracked.
What I need to remember that the summary will not contain. The tone. The fact that someone was clearly unhappy. The thing said just before the call ended. The reason behind a decision, which never survives into a bullet list.
Ninety seconds. That produces the artefact, and now the transcript is a genuinely useful thing to have behind it — you have a short document with the meaning in it, and a long document you can search when you need the exact words.
Attention, and what it costs to stop paying it
The uncomfortable part.
Meetings where everyone knows it is being recorded are measurably less attentive meetings. People check messages, half-listen, and rely on the summary. And the summary does a reasonable job of the content, which means nothing appears to be lost — except the thinking that used to happen during the meeting because people were engaged with it.
That is a real cost and it is invisible in the artefacts. The meeting produced a good transcript and a worse outcome.
The version that works is deliberate: the recorder handles the record, and you use the attention it freed up on the discussion rather than on your messages. That is a decision to make on purpose, because the default is not that.
Consent and where the recording lives
Two things worth thinking about once rather than never.
Recording people has legal and social requirements. Rules on consent vary by jurisdiction, sometimes significantly, and this is genuinely worth checking for the places you work rather than assuming the tool has handled it. Socially, a bot joining unannounced changes what people say, which is a cost to the meeting whether or not it is lawful.
A transcript is one of the most sensitive documents an organisation produces. It is a verbatim record of what people said, including the parts they would not have written down — offhand remarks, complaints, half-formed opinions about colleagues. It sits on the transcription vendor's infrastructure, and it is discoverable, breachable, and reviewable by whoever has admin access.
If it would matter for a conversation to be read back verbatim, that is a reason to take notes rather than record. The three-line artefact contains what was decided. The transcript contains what everyone said, which is a different and much larger liability.
Where each thing belongs
Transcripts stay in the meeting tool. They are bulky, they are the organisation's record, and they belong where they were produced.
Your three lines go in your own notes. In the same collection as everything else you write, so a decision from March turns up when you search for the project. They leave with you, they contain no verbatim record of anyone else, and they are short enough to actually reread.
This split is the same one that applies to work notes generally: shared records go in the shared system, your own understanding goes somewhere that is yours.
The honest version
If you run a lot of meetings, use a transcription tool. Do not stop — the record is genuinely better than what you would write, the search is useful, and being able to check what someone actually said has settled more disagreements than any minutes.
Just do not mistake it for notes. It is a recording. The notes are the ninety seconds afterwards where you decide what it meant, and that step has not been automated because it is not a summarisation problem. It is a what mattered to me problem, and nothing else knows the answer.
Cyanote has no transcription and no AI summarisation — it is the place the three lines go. Meeting and standup templates in the / menu, tasks that take an owner and a date out of a typed sentence, [[ links so a decision attaches to the project and the person, and full-text search across every meeting note you have written. It is a local database on your own Mac, which is a reasonable place for the short document containing what things meant, whatever your organisation does with the long one.