AI note-taker privacy: where your meetings really go

An AI note-taker turns your most candid professional conversations into files — audio, transcripts, summaries — that live on somebody's infrastructure under somebody's policy. Before you compare summary templates, compare answers to the questions on this page. We wrote it because vendor trust pages are where enthusiasm goes to hide.

The five questions that sort the market

1. Where does processing happen?

Most note-takers are cloud services: audio leaves your machine, gets transcribed and summarized on vendor servers, and comes back as notes. The documented exception on our board is Krisp, which runs English transcription on-device — raw audio can become text without leaving your computer — with additional languages processed server-side, and on-device options expanded at Enterprise tier. If "audio never leaves the building" is a hard requirement, that architectural difference is your starting point, and you should verify the current details against Krisp's own documentation for your exact plan and languages.

2. How long is everything kept?

Recordings and transcripts persist until something deletes them. The operative questions: what's the default retention, can admins set shorter policies, and does deleting in the app actually purge backend copies within a stated window? Fireflies documents custom data retention at Enterprise; other vendors put retention controls at various tiers. If a vendor's answer to "how do I delete a meeting everywhere?" takes more than a paragraph, notice that.

3. Is your audio training their models?

The question everyone asks after signing up. Vendors in this category generally state that customer meeting content isn't used to train shared models, or offer an opt-out — but the wording moves, differs by tier, and occasionally differs between the marketing page and the privacy policy. Read the privacy policy's actual sentence, check for an admin toggle, and if the answer matters to you, get it in writing from sales. We deliberately don't reproduce each vendor's current sentence here, because these clauses are edited too often for our word to be worth anything — always confirm at the source, today.

4. Who inside your org can see what?

The under-asked one. A searchable archive of every meeting is an insider-risk surface: the intern can now query what the executive team said, unless workspace permissions say otherwise. Look for workspace roles, private-by-default meeting settings, channel or folder scoping (Fireflies documents private channels at Business tier), and audit logs at team tiers. Then decide your policy — who can search old transcripts is a governance question no vendor answers for you; our consent guide covers the human half.

5. What certifications and controls back the claims?

Look for the standard artifacts — SOC 2 reports, GDPR/DPA terms, SSO/SCIM at enterprise tiers, encryption statements, and for health-adjacent work, whether the vendor will actually sign a BAA (Krisp advertises HIPAA compliance options at Enterprise; verify scope with them directly). A trust page listing badges is a start; the report behind the badge is the substance, and vendors provide it under NDA to serious buyers.

Practical hygiene, whatever you pick

• Default to less capture. Record the meetings whose minutes matter, not everything with a calendar entry. The cheapest data breach is the recording that never existed.
• Keep sensitive meetings out of scope. HR conversations, legal strategy, anything with health details — off the note-taker by policy, per the consent guide's off-the-record protocol.
• Set retention short and forget-friendly. Ninety days of transcripts answers "what did we decide"; seven years of them answers a subpoena.
• Route summaries, not raw transcripts. Summaries travel to channels and CRMs; transcripts carry every aside and misfire. Share the digest, restrict the verbatim.
• Use the workspace controls you paid for. SSO, roles, private meetings — team tiers include them because rollouts without them go feral.
• Re-check clauses on renewal. Training and retention language changes. Put "re-read the privacy policy" in the renewal checklist next to the price.

How this shapes our scores

Privacy controls carry 10% weight in our rubric — enough to reward documented architecture, not enough to pretend we've audited anyone. Krisp's 8 reflects the on-device English transcription path; the cloud-only tools sit at 5–6 with differences for documented controls. If your weighting would be higher, our script is public — tools/score.py — re-run it with your numbers.

Fair questions