The most tedious part of a call comes after the call
An hour on the phone with a client is only half the trouble. The real work starts afterward: sit down, listen to the recording again, fish out what was agreed, and put it into text. A one-hour call easily eats up another hour of fiddling. And that whole time I'm not working on the project - I'm working on the recording.
For a long time it felt like an unavoidable tax on every call. Then I noticed that the most useful way AI fits into my day isn't where it "writes for me" at all, but in this quiet routine of re-listening and transcribing.
How I have it set up
The setup turned out to be simple. I record the call, run the recording through speech recognition right on my own machine, and get text out. Then AI assembles a short summary from the transcript: what we agreed on, what we decided, and what happens next. I review the draft myself - the model can miss a caveat or state too confidently something that was only being discussed - and only then send the summary to the client.
You may only record a call with an explicit heads-up and the other person's consent. The first thing I do on a call is say that I'm recording it and explain why - to put together a summary of what we agree on. No consent, no recording. This isn't a formality, it's the condition without which the whole method has no right to work.
The key thing here is that there's nothing left to re-listen to. An hour of conversation fits into three paragraphs you can read with your eyes in a minute. And I'm not the only one reading them: the client gets a clear "here's what we agreed on" and confirms it or corrects it.
What it actually gave me
There are noticeably fewer misunderstandings. It used to be that a discrepancy - "wait, I thought we said something different" - would surface a week later, when redoing the work is already expensive. Now it surfaces the same day, at the level of the text, where a fix is a single line.
My time freed up too. The hours that went into rewinding recordings came back into the work, and on the call itself it became easier to actually listen to the person instead of trying to write the perfect notes at the same time. What's left to review is a short piece of text, not the whole recording from start to finish.
It's especially valuable that the recognition itself runs locally, on my machine: I don't have to upload the audio to a transcription service. But that doesn't automatically make the whole process local. If an external model is used for the summary, the transcript is handed to it - which means that before the call I need to understand whether that's acceptable for a given client and what data can be sent. A fully local model removes that question too.
For now it's a habit stitched together from separate manual steps. But more and more I want the recording, the transcription, the summary review, and sending it to the client to live in one clear flow right inside the editor where I already spend the whole day. It feels like this is worth building into a separate tool.
AI helps most not when it writes for you, but when it takes away the drudgery you were putting off the actual work to avoid.
