I take bad notes. I always have. So when I started seeing meeting-transcription tools everywhere, the pitch landed: let the machine do the part I am demonstrably bad at.
I gave it a month. Every meeting, nothing written by hand, tool running.
Here is what actually happened.
The transcription is basically solved
I expected this to be the weak part. It was not, and it was not close.
Across four weeks of real meetings — accented English, people talking over each other, a large amount of internal jargon and at least two participants who mumble — the raw transcript was accurate enough that I stopped spot-checking it after the first week. That is a genuinely different situation from two or three years ago.
The summaries were better than my own notes, too, and I want to be honest about why: not because the tool is clever, but because it does not get bored. My notes are detailed for ten minutes and then become three words and an arrow. The tool’s attention is flat across an hour. For “what was discussed and what was decided,” it beat me comfortably.
Then I looked at the action items
This is where it fell over, and the failure is specific rather than general.
AI note-takers capture tasks that were stated as tasks. Someone says “I’ll send that over by Thursday” and it lands, correctly attributed, with the date.
They miss tasks that were implied, which in my meetings is most of them. When someone says “that’s going to be a problem for the migration timeline,” work has just been created for somebody. Nobody assigned it. Nobody named it. Everyone in the room understood it. No tool I used caught a single instance of this.
The result is worse than an incomplete list, because it does not look incomplete. You get a tidy summary with a section labelled Action Items, and the things missing from it are precisely the ones that required understanding the conversation rather than hearing it.
The thing nobody warns you about
About three weeks in I noticed I was walking into follow-up conversations with much less in my head than usual.
Not that I had forgotten things — I could search the transcript for anything. But searching is not the same as knowing, and the corridor conversation where someone asks what you thought about the pricing question does not pause while you search.
This is not a novel observation; there is research going back over a decade on note-taking as an encoding activity rather than a storage one. Deciding what is worth writing down is doing cognitive work whether or not you ever read the note again. I had been treating notes as a filing problem, and it turns out they were also doing something to my memory that I had not been paying for consciously.
Removing it entirely was the mistake. The tool was fine.
The awkward part
A bot joining a call changes the room. People are more careful, less speculative, less likely to think out loud — and thinking out loud is where a lot of the value in a meeting actually is.
There is also the legal question, which the marketing pages step around. Recording rules vary by jurisdiction and some require every party to consent rather than just one. Employers frequently have their own policy on top. Telling a client mid-call that you have been recording is a conversation you want to have once, in advance, rather than improvised.
Sort that out before it becomes a habit rather than after.
Where I landed
I kept the tool and changed how I use it.
It runs for the transcript and the summary, which it does better than I do. And I write three lines by hand during the meeting — not a record, just the things I want to still be carrying afterwards. It takes almost no effort and it restored the part I had accidentally given away.
The pitch for these tools is that they replace note-taking. What they actually replace is transcription, which is the part of note-taking that was never the point.
