You can ask for it back. You cannot un-send it.

To delete your data from a cloud AI notetaker: revoke its calendar and meeting access, delete every recording and transcript in the app, then submit a formal data-deletion request to reach the copies you can’t see — backups, subprocessors, and training sets. Here is the full procedure, and the honest limit on what it can achieve.

The five-step deletion procedure

The in-app “delete” button removes what is visible to you. Getting to everything the vendor holds takes five steps, in order:

1. Cut off new capture first. Revoke the notetaker’s access to your calendar and conferencing accounts (Google/Microsoft OAuth, Zoom/Meet/Teams app grants). Otherwise it keeps ingesting new meetings while you delete the old ones.
2. Delete recordings and transcripts in-app. Remove every meeting, then empty any trash or archive. Note that this typically clears the UI copy only.
3. Delete derived artifacts. Summaries, action items, CRM syncs, and shared links are separate objects. A deleted meeting can leave its summary live in a connected Slack channel or CRM record.
4. Submit a formal erasure request. The in-app controls do not reach backups, subprocessors, or training data. A written request under GDPR Article 17 or CCPA does — template below.
5. Get written confirmation. Ask for confirmation that deletion propagated to backups and subprocessors, with a date. Without it, you have deleted your view of the data, not the data.

The three copies deletion usually misses

Even a diligent pass leaves three copies behind, because they live outside the interface you were given:

• Backups. Deletion of the live record rarely purges encrypted backups immediately; they age out on the provider’s own schedule, often weeks later.
• Subprocessors. If the tool streamed your audio to third-party ASR or LLM vendors, your data reached servers you never had an account on. Your deletion request has to be relayed to them — which is why step 5 asks for the list.
• Training sets and model weights. If a recording was used to train a speech or summarization model, deleting the source file does not remove what the model already learned. This is the copy that cannot be meaningfully clawed back.

A data-deletion request you can send today

Send this to the vendor’s privacy or support address. It invokes the legal rights that reach past the in-app controls:

Subject: Data deletion request (GDPR Art. 17 / CCPA)

I am exercising my right to erasure. Please, for the account and any
meetings associated with my email address:

  1. Delete all audio recordings and derived audio representations.
  2. Delete all transcripts, summaries, and meeting metadata.
  3. Delete any voiceprint or biometric identifier derived from my voice.
  4. Confirm whether any of the above was used to train models, and
     remove it from training sets and future model updates.
  5. Instruct all subprocessors (ASR / LLM / storage vendors) to do
     the same, and list them.
  6. Confirm deletion in writing, including from backups, within 30 days.

Item 3 matters more than it looks: several notetakers have faced biometric-privacy litigation over voiceprints, and a voiceprint is a separate artifact from the recording it was derived from. Item 4 is the one vendors most often answer vaguely — a clean workflow gives you a straight yes or no.

Why metadata outlives the recording

Suppose deletion works perfectly. The vendor can still retain meeting metadata — who met, when, for how long, with what title — because it is often treated as operational data rather than content. For a privileged or clinical relationship, the metadata alone can be sensitive: the fact that a specific client met a specific attorney on a specific date is sometimes the confidential part. Deletion addresses the transcript; it rarely addresses the ledger of who spoke to whom.

The alternative: nothing to delete

Deletion is remediation. It is the work you do because audio was captured, sent, and stored, and now you would like it gone. An on-device notetaker removes the need for the entire procedure by never producing the artifact in the first place.

Fazit holds call audio in a RAM ring buffer, transcribes it with an on-device model, and writes the note to your own vault as Markdown. There is no recording to delete, no subprocessor to send an erasure request to, and no training set your voice could have landed in — because the audio never left the machine. The mechanism behind that guarantee is in Why “Never Records” Is Not Marketing — It Is an Invariant, and the architecture-level comparison to cloud tools is in On-Device vs. Cloud AI Notetakers.

Deletion is a promise a vendor makes about your data. “Never recorded” is a property of a system that never held it. If you have reached the point of drafting erasure requests, that is the distinction worth acting on.