Where each AI notetaker actually stores your data
Feature tables rarely answer the question a compliance review asks. This one does.
Every AI notetaker says it takes privacy seriously. That sentence costs nothing to write. The question that actually determines your exposure is narrower: does the audio leave the building, and if so, whose building does it land in?
We make one of these tools, so read the Oats column with appropriate suspicion. Everything about the others is drawn from their own public documentation, linked at the bottom — check it, because vendors change terms.
The comparison
| Otter | Fireflies | Granola | Oats | |
|---|---|---|---|---|
| Bot joins call | Yes | Yes | No | No |
| Audio processed | Cloud | Cloud | Cloud | Your device |
| Transcript stored | Cloud | Cloud | Cloud | Your device |
| Account required | Yes | Yes | Yes | No |
| Trains on your data | Contested in court | States it does not | Opt-out on Enterprise | Structurally cannot |
| Subject to subpoena | Yes | Yes | Yes | Only you are |
| Works offline | No | No | No | Yes |
The distinction that matters: policy vs architecture
Three of the four rows above are promises. Fireflies states it does not train on customer data. Granola offers training opt-out on its Enterprise tier. These are real commitments and we have no reason to doubt them.
But a promise is a choice a company makes, and choices get revisited — at acquisition, at the next funding round, when a new general counsel arrives. Promises also don't cover events the company doesn't control: a breach, a misconfiguration, a lawful order to produce records.
An architectural claim is different. If the audio never reaches a server, there is no dataset to repurpose, no bucket to leak, and nothing for a third party to be compelled to produce. That's a weaker-sounding guarantee that happens to be much harder to break.
The flip side, stated plainly: local-first moves the risk onto you. An unencrypted laptop in an airport is a worse custodian than a company with a real security team. Turn on FileVault. Back up. Local-first is not a security strategy by itself — it removes one class of risk and hands you another.
The questions a compliance review actually asks
If you're a therapist, lawyer, clinician, or consultant, the form in front of you usually wants:
- Who are your sub-processors? Cloud notetakers are one, and typically bring their own model vendor as a second. Local processing has none to list.
- Where is the data at rest? “On the practitioner's device, encrypted at rest via FileVault” is a sentence most reviews accept.
- What's the retention period? With a cloud tool this is theirs to set. Locally, it's whenever you delete the file.
- Can it be produced on request? Someone will eventually ask for your records. Knowing whether a vendor also holds a copy matters.
Choosing sensibly
Cloud notetakers are the right answer for plenty of teams. Shared workspaces, CRM sync, and recording meetings you don't attend are real capabilities that local-only tools structurally cannot offer.
The trade is worth making consciously rather than by default. If a conversation would be awkward to explain having uploaded — a client session, a diagnosis, a deposition prep, an unannounced acquisition — that's the signal to keep it on the machine it happened on.
Sources: Granola pricing, Otter pricing, Fireflies pricing. Accurate as of July 2026 to the best of our knowledge; corrections welcome via GitHub.