There are only three things a notetaker can do with your audio.

An on-device AI notetaker transcribes your call on your own machine and never sends the audio anywhere. A cloud notetaker ships that audio to a vendor’s servers to transcribe it. Everything else — summary quality, integrations, price — converges every quarter. This one distinction does not, because it decides who else can ever hold your client’s voice.

The two architectures, defined

Cloud transcription captures your call audio and transmits it to the vendor’s infrastructure — often a bot that joins the meeting as a visible participant — where it is stored, transcribed, summarized, and (per policy) retained or deleted on a schedule. The audio leaves the room.

On-device transcription captures the audio your own machine already hears and runs the speech model locally, on the device’s own neural hardware. No audio is transmitted; the transcript is produced where the audio was captured. Nothing leaves the room.

The marketing language blurs this on purpose. “Local-first,” “private by design,” and “we delete it after” all sound adjacent, but they describe materially different data paths. The useful frame is not the adjective a vendor picked — it is which of three states your audio actually ends up in.

The three states audio can be in

STATE                  WHAT IT MEANS                         WHO LIVES HERE
Created, then deleted  Audio written to a vendor disk,       Cloud recorders with a bot
                       then deleted on a retention timer     (Fathom, Otter, Fireflies, tl;dv)
Sent, then deleted     Audio streamed to ASR/LLM vendors,    Bot-free but cloud-transcribed
                       raw file deleted after transcription  (Granola)
Never created          Audio held in RAM, consumed by an     On-device transcription
                       on-device model, never written        (Fazit)

Almost every privacy debate in this category is really an argument about which of these three states a given tool lives in.

• Created, then deleted. The classic cloud recorder. A file exists on a vendor’s disk for some window. During that window it can be subpoenaed, breached, or used to train a model. Deletion reduces the window; it does not change the fact that the file existed.
• Sent, then deleted. The bot-free-but-cloud-transcribed middle ground. Better — no bot, shorter retention — but your conversation still transits the public internet to third-party ASR and LLM providers before anything is deleted.
• Never created. On-device transcription. The audio is held in volatile memory, consumed by a local model, and discarded when the process exits. There is no file, no transfer, and no third party holding it.

Why “we delete it” is a promise, not a property

“We delete your audio after transcription” is the most common privacy assurance in the category, and it is worth reading precisely. It is a statement about what a vendor does after your audio arrives on their server — not a guarantee that it never arrived. Deletion is a behavior you have to trust: it depends on the vendor’s retention job running, on their subprocessors honoring the same policy, and on no backup, log, or training snapshot having captured the data in the meantime.

“Never created” is not a behavior you have to trust. It is a property of the data path. If there is no code path that writes audio to disk or sends it over the network, there is nothing to delete, nothing to leak, and nothing to disclose in a subprocessor list. A property is stronger than a promise because it does not depend on anyone keeping their word.

How to tell which one you’re actually using

Vendor copy will not reliably tell you. These five questions will. Ask any tool — including this one — and the honest answers place it in one of the three states above.

1. Does raw audio leave the device that captured it? If yes, you are in a cloud state regardless of the “private” branding.
2. Is there a subprocessor list? A tool that transcribes on-device has no ASR or LLM vendor to name. If the list exists, the audio is going somewhere.
3. Does a bot join the call? A visible participant is a recorder by definition, and it captures everyone — not just you.
4. What is the data-retention window, and what is exempt from it? Backups and training sets are frequently carved out of the deletion promise.
5. Can the vendor produce your audio in response to a subpoena tomorrow? If the honest answer is “only until the retention timer fires,” the audio exists today.

A tool that answers no / none / no / n/a / no down that list is doing on-device transcription. Anything else is some flavor of cloud, and the only real question left is how short the window is.

Why this matters more in regulated work

For a casual internal standup, “deleted-after” is usually fine. The calculus changes the moment the conversation is privileged, clinical, or fiduciary. In all-party-consent jurisdictions, a bot that records everyone can create liability before a word is transcribed. Under GDPR, sending EU call audio to a US vendor is a cross-border transfer that needs a lawful mechanism and a transfer-impact assessment. In law, medicine, and finance, the existence of a recording is itself discoverable.

On-device transcription sidesteps the entire chain: no transfer to assess, no recording to discover, no subprocessor to bind. That is not a marginal compliance advantage — for a confidentiality-bound professional it is often the difference between a defensible workflow and an indefensible one. The legal reasoning behind “never created,” including the German §201 StGB distinction between audio that was never written and audio that was written then deleted, is laid out in Why “Never Records” Is Not Marketing — It Is an Invariant.

Where Fazit sits

Fazit is the “never created” row. Audio is captured into a fixed-size RAM ring buffer, a local model on Apple Silicon consumes a snapshot, and the note is written to your vault as Markdown. The audio never leaves the machine — so there is no subprocessor to list, no transfer to disclose, and no deletion step to take on trust. Every note it writes carries audio_retained: false in its frontmatter, because that line describes the execution path rather than a policy.

That design is a trade, not a free lunch: it is Apple Silicon only, it will not send a proxy bot to record a meeting you didn’t attend, and its integration surface is deliberately smaller than a cloud CRM autopilot’s. If you want the tool-by-tool version of this argument — where each of Fathom, Otter, Fireflies, tl;dv, and Granola actually lands — read Fazit vs. the Cloud Notetakers: Where Does the Audio Actually Go?.