Start with the honest part, because nobody else will: there is no independent measurement of how accurate these tools are.
Search for a 2026 accuracy benchmark and you will find a dozen tidy tables ranking Otter against Fireflies against Granola. Check the domain on each one and you will find it belongs to a competing product. None publishes a test corpus, a methodology, or anything reproducible. They are marketing, not measurement. The nearest neutral yardstick, the Open ASR Leaderboard, benchmarks open speech-recognition models on public datasets — it does not test any of these products.
Every “95% accurate” claim in this category is self-reported and unaudited, against undisclosed audio. So this comparison is about the things that are verifiable: price, limits, where your recordings live, and whether a bot joins your call.
What the free tiers really give you
| Tool | Free tier | Paid from |
|---|---|---|
| Otter.ai | 300 min/month, 30 min per conversation, 3 lifetime imports, only the 25 most recent conversations kept | $8.33/user/mo annual |
| Fireflies.ai | 400 min per team (not per user), 20 AI credits | $10/seat/mo annual |
| Granola | Unlimited meetings, but only the last 30 days visible; training opt-out included | $14/user/mo |
| Fathom | Unlimited recording, storage and transcription in 38 languages — capped at 5 advanced AI summaries a month | $16/mo annual |
| Zoom (free plan) | 3 meeting summaries, 3 AI note-taking uses, 20 AI queries a month | Bundled with paid Zoom plans |
Fathom is the outlier and deserves the note: its free tier caps intelligence, not volume. Unlimited recording and storage, then basic templates only after five smart summaries — no action items, no follow-up emails, no “Ask Fathom”. That limit is in the help centre, not on the pricing page, which is worth knowing before you build a habit on it.
Where your meetings live, and who trains on them
| Tool | Storage | Trains on your data? |
|---|---|---|
| Otter | AWS, United States | Yes — trains on de-identified audio and on transcriptions “which may contain Personal Information”. No clear opt-out is described. |
| Fireflies | US servers “and other countries” | No, stated absolutely, with vendors contractually prohibited too. |
| Granola | US AWS VPC — and no audio is stored at all | Partly — third parties barred; Granola trains on anonymised data with a user opt-out. |
| Fathom | United States, provider not disclosed | Partly — de-identified content improves internal models, opt-out in settings. |
| Zoom | US data centres | No — “Zoom does not use any customer audio, video, chat… to train Zoom’s or its third-party AI models.” |
Bot joins, or records locally — the choice that matters most
Otter, Fireflies and Fathom send a bot into the meeting. Granola does not — it captures your microphone and system audio on your own machine, transcribes in real time, and stores no audio. Zoom, Google Meet and Teams handle it server-side, visible in the UI rather than as a participant.
Three consequences follow, and they are bigger than any feature list:
- Consent. A visible bot is de facto notice to the room. Local recording is not. In a two-party-consent state, the person who started the recording carries that risk personally.
- Control. A bot has to be admitted, so hosts and IT can block it. Local capture is invisible to the host entirely.
- Retention. Bot tools keep audio server-side by default. If there is no recording, there is nothing to leak, retain or subpoena.
This is not hypothetical. In August 2025 a putative class action, Brewer v. Otter.ai, was filed in the Northern District of California alleging Otter “deceptively and surreptitiously” recorded conversations without all-party consent and used them for AI training, under state and federal wiretap law. Otter’s position is that users authorise training use and that audio is de-identified first; the complaint disputes whether that de-identification is effective. Nothing has been proven and the case is unresolved — but it is a live reminder that the person who invites the bot is the one on the hook.
What is built into the platform you already pay for
- Zoom. AI summaries are bundled with paid Workplace plans at no extra cost. Free accounts get 3 summaries a month. ZoomMate, launched 1 June 2026, is a separate paid agentic layer — Zoom’s own community guide says it starts at $20/user/month while its product page implies $15–16; those two official sources disagree, so get a live quote.
- Google Meet. “Take notes for me” needs an eligible Workspace edition or a Google AI plan. Eight languages, one per meeting, best between 15 minutes and 8 hours. Notes land in the organiser’s Drive. As of August 2026 it also works for in-person meetings.
- Microsoft Teams. Intelligent recap comes with Teams Premium at $10/user/month billed annually, on top of a paid Teams licence. Microsoft 365 Copilot is a separate licence and Microsoft does not publish its price on the Teams Premium page — we are not going to repeat the number everyone quotes without a source.
If your team is already on a paid Zoom or Workspace plan, try the built-in option for two weeks before paying for a fourth tool. It is frequently good enough, and it adds no new company to your data map.
Picking one
- Privacy-first, no bot: Granola. No audio stored, opt-out on the free plan.
- Most free value: Fathom — unlimited recording and storage, five smart summaries a month.
- Clearest no-training stance: Fireflies, which states it flatly and extends it to vendors.
- Already paying for Zoom or Workspace: use what you have first.
- Otter: the best-known name, the tightest free tier, the only one that trains on transcriptions with no described opt-out, and the one currently in court over it.
One last thing no tool fixes: notes nobody reads. If the summaries pile up unread, the problem is downstream of transcription — it is a review problem.
Leave a Reply