What happens to your audio when you use an AI notetaker?

The usual path your audio takes

With most AI notetakers, the flow is the same under the hood. A bot joins your call as a participant. The audio is streamed or uploaded to the company’s servers. There it is transcribed, summarised, and stored in your account, where your team can search it later. The convenience is real, and so is the fact that your meeting now lives on someone else’s infrastructure.

A few details are worth knowing. The bot is usually visible in the participant list, so the other people can tell a notetaker is present. The recording and transcript are often retained until you delete them, and the retention period varies by provider. Some services use recordings to improve their models unless you opt out, which is the kind of thing that sits in the terms rather than the marketing.

When that is fine, and when it is not

For a routine internal sync, a cloud archive your team can search is useful, and the privacy trade is minor. The calculation changes when the meeting is sensitive. A one-to-one about someone’s performance, a call under NDA, anything involving a customer’s private data: those are the meetings where uploading the audio to a third party is a real risk, not a theoretical one.

The point is not that cloud notetakers are bad. It is that where your audio goes should be a decision you make on purpose, not a default you never noticed.

The on-device alternative

The other option is to keep the whole thing on your own machine. An on-device tool records the meeting locally and transcribes it on your Mac, so the audio is never uploaded. No bot joins the call, because the tool taps the audio your Mac is already playing instead of dialling in.

If you go this route, two things are worth checking. First, that the app is sandboxed and declares no data collection, so you know the audio is not quietly leaving by another door. Second, that transcription really is local, not a cloud call dressed up as a local feature. A tool that only touches the network to download its speech model once, and works with Wi-Fi off after that, is doing it properly.