They removed the bot. They did not remove the server.

2026 is the year the meeting bot died. Granola made bot-free capture famous; Krisp built it into the audio driver; Circleback added a desktop app; then Fireflies and Fathom — the two biggest bot companies — shipped bot-free modes of their own. The industry heard the feedback: nobody wants a third participant named “Notetaker” staring at their client. But somewhere in the wave, “bot-free” started being sold as if it meant private. It does not. Botless describes what your meeting looks like. It says nothing about where your audio goes.

Why everyone went botless

The bot was a social problem before it was a privacy problem. A visible recorder changes how people talk — surveys of professionals consistently find large majorities self-censor when a notetaker is in the room. Clients question it; some organizations now block third-party bots from joining meetings at all; bar associations warn lawyers about them. The bot became a liability you could see.

So the category removed it. Capture moved from a fake participant in the meeting to the desktop: the app records your microphone and system audio directly, and no one on the call sees anything. Socially, this is a real improvement — it is the reason this product category is usable in front of clients at all.

What botless does not change

Follow the audio after capture and the botless apps look exactly like the bot apps: the recording is streamed to the vendor’s cloud, transcribed by third-party speech services, summarized by cloud LLMs, and the transcript is stored in the vendor’s database under your account. The bot was the visible tip of the pipeline. The pipeline is still there.

                    BOT IN THE     AUDIO LEAVES      TRANSCRIPT STORED
                    MEETING?       YOUR DEVICE?      ON VENDOR SERVERS?
Otter               yes            yes               yes
Fireflies (bot)     yes            yes               yes
Fireflies (app)     no             yes               yes
Fathom (bot-free)   no             yes               yes
Granola             no             yes               yes
Circleback (app)    no             yes               yes
Krisp               no             partly*           yes
Fazit               no             no                no — files in your vault

* Krisp offers on-device transcription for English; AI notes are cloud.

To be fair to the field: these are mostly honest, well-run clouds — SOC 2 audits, no-training commitments, EU hosting in some cases, and Krisp genuinely runs English transcription on-device. If your threat model is “I trust vendors, I just hate the bot,” a botless cloud notetaker is a fine choice. But a security review does not ask what your meeting looks like. It asks where the data goes — the framework from On-Device vs. Cloud AI Notetakers. And on that question, botless-cloud and bot-cloud are the same answer.

The tell: what happens to the audio

Every cloud notetaker’s privacy page eventually says some version of “audio is deleted after transcription.” Read that sentence carefully: it concedes the audio was there — on their servers, in transit through their subprocessors — and asks you to trust the deletion. For a casual standup, fine. For a privileged legal call, a therapy-adjacent session, or a fiduciary conversation, the record having existed on third-party infrastructure is itself the exposure: it is subject to subpoena, breach, retention mistakes, and policy changes. The consent-law side of that risk is covered in Is It Legal to Record Client Calls?

The third step: botless, then serverless

Removing the bot fixed the meeting. The next step removes the server from the data path entirely. Fazit runs the whole pipeline on your Mac: audio is captured into memory, transcribed by an on-device speech model, summarized by a local LLM, and written as a Markdown file into your Obsidian vault. The audio is never written to disk and never leaves the machine — there is no “deleted after transcription” promise because there is nothing to delete and nowhere it went. No account holds your transcripts; a folder of files you own does.

The honest trade: this only works if the models run on your hardware, so Fazit is Apple Silicon-only, built for 1:1 client calls rather than recording meetings you did not attend — and it is early software. The architecture that makes the guarantee enforceable rather than promised is documented in Why “Never Records” Is Not Marketing.