I've been the one who takes the minutes in the team for long enough to admit the dirty secret: the notes are the least useful thing in the room, because the person writing them usually isn't listening. The whole "AI meeting notes vs manual notes" debate usually gets framed like a budget argument or a productivity flex. It's neither. It's an attention argument, and it's the one I keep losing to my own pen.
When you're the one scribbling, you're not really hearing — you're queuing up what the last speaker said while the current one keeps going. One-on-ones are survivable. A status sync with eight people in it is where manual notes fall apart, and that's exactly the meeting where the decisions actually happen. People talk over each other, someone commits to a deadline in half a sentence, and the writer's hand is two topics behind.
A transcript doesn't fix it either. I've read plenty. A transcript is a wall of words nobody re-reads, and you still have to dig the decisions and the action items out yourself — which is the part that matters and the part that always got dropped.
AI meeting notes vs manual notes: it's an attention split, not a cost thing
What changed for me is recording the meeting and letting a machine do the dreary work — but on the machine in front of me, not on some cloud. I don't want our vendor conversations sitting in someone's training set, and I suspect the meetings most worth recording are the ones you'd least want parked on a stranger's server.
So here's the pipeline: the app grabs my mic and the system audio (so Teams, Zoom, Meet, or an in-person table all work), hands it to Whisper for transcription, then a small local LLM running in LM Studio writes the minutes. The output is minutes, not a transcript: a summary, the key decisions, an action items section with owners and dates where anyone bothered to name them, and the open questions saved for the next call. Then it exports to Markdown, text, or PDF, so the thing lands in email, in the ticket, or just as a file on my own disk.
Runs on a laptop CPU fine; a GPU just makes the transcription quicker. Needs Windows 10 or 11, Python 3.10+, and a free copy of LM Studio with any small model loaded. That's the whole setup.
The payoff I didn't expect was attention. I walk out of a meeting having actually said something instead of having written something. I catch when a question stalls. I notice who went quiet. I stopped asking people to spell their action items out again.
It's not magic. Whisper's transcript deserves a glance, and it'll happily time-stamp a "we'll circle back" as an action if you don't correct it. But a thirty-second pass over the action items beats thirty minutes of writing minutes, and it beats the whole "we'll update the notes later" ritual dead — the minutes are done when the call ends.
If the "take your own minutes locally" idea sounds like your problem too — that's the tool I built, because the note apps kept sending my recordings somewhere I couldn't control. It's called Local Meeting Notes: a one-time $9, no subscription, nothing ever uploaded. If you'd rather keep your meetings on your own machine, it's worth a look at https://symshah.gumroad.com/l/xgoypx
FAQ
Can it run without a GPU? Yes. Whisper transcription works on CPU, just slower. A GPU only speeds up the transcribing step.
Does it work with Teams and Zoom? It records both your microphone and the system audio, so any call that plays out of your speakers gets captured, plus in-person meetings on the same Windows machine.
Is anything uploaded? No. Recording, transcription, and the minutes all stay on your PC. There's no account and no cloud — that's the point.
Top comments (0)