How to choose an AI note taker

An AI note taker joins or records your meetings and gives you a transcript, a summary and a list of action items. They look similar on a pricing page. In practice they differ in five places: how accurate the words are, who said what, how the tool gets into the meeting, what happens to your data, and which languages it really handles.

This guide explains what to check. It does not rank anything. We have not finished testing any note taker yet, and every tool on our ranking page is marked NOT YET TESTED until we have.

First lab results: Lab run #1 includes a synthetic STT-01-TTS transcript scored with Whisper-normaliser WER on Cloudflare Workers AI Whisper models. Consumer meeting-note apps (Otter, Fireflies, Fathom, Granola, Read AI, tl;dv, Zoom) are still NOT YET TESTED. The AI meeting note taker ranking stays unranked until those app tests ship.

1. Transcript accuracy: the words themselves

Everything downstream depends on this: summaries, action items, search. A summary built on a transcript that heard "$89" instead of "$79" is confidently wrong.

What to check yourself, on a free plan:

• Record a 10-minute call with names, numbers and product terms in it. Then read the transcript next to the audio. Count the mistakes on the words that matter (prices, dates, names), not on "um" and "uh".
• Try your real conditions: a laptop mic, one person on a phone, someone with an accent.
• Look at how numbers are written. "Thirteen twelve" and "1,312" may be the same thing, or a mistake.

How we measure it: word error rate (WER) on a scripted call where we know every word that was said, plus a separate check on the facts and numbers we planted in the script. See how we test.

2. Speaker labels ("diarization")

"Someone agreed to ship by Friday" is useless. You need who agreed. Speaker labelling is a separate problem from transcription, and it gets harder with more people, crosstalk and similar voices.

Check:

• Does the tool name speakers from the meeting platform, or does it guess "Speaker 1, Speaker 2"?
• With 4+ people, are action items assigned to the right person?
• When two people talk at once, does one of them disappear?

3. Bots vs bot-free capture

There are two ways a note taker gets your audio:

• A bot joins the call as a visible participant. Everyone sees it. That is good for transparency, but some hosts block bots, and the bot depends on the meeting platform's rules.
• Bot-free capture records on your own computer (a desktop app or browser extension). Nothing joins the call, but you are responsible for telling people they are being recorded.

Things that changed recently:

• Zoom now requires authorization for bots. Zoom's developer documentation says bots that join meetings outside their own account must use an "On Behalf Of" token from an authorized user who is in the meeting. Enforcement started on 2 March 2026 (Zoom developer FAQ, checked 2026-10-08). One note-taker vendor tells its users that Zoom enforces this for its integration from 14 October 2026 (Fellow help center, checked 2026-10-08). In practice, a bot may leave when the person who authorized it leaves.
• Bot-free options are often limited by platform. For example, Fathom's pricing page lists a "choice of bot-free (in beta) or bot capture" (fathom.ai/pricing, checked 2026-10-08). Granola's help center describes desktop apps for macOS and Windows (Granola docs, checked 2026-10-08). If you're on Linux or a locked-down work laptop, check before you commit.

4. Privacy: where your meetings go

Meetings contain salaries, customer names and health details. For this category, privacy carries 15% of the score in our methodology. Read the vendor's privacy policy and trust page, not the marketing page. Check these items:

Different vendors make different choices here. That is why we quote the exact sentence and link it on each tool page, rather than summarise it as "secure".

5. Hebrew and other languages

"Supports 100 languages" can mean "you can pick it in a menu". It doesn't tell you the transcript is usable. Check two things:

1. Is your language on the official list at all? Lists vary a lot. Otter's help center says it "currently only supports English, Spanish, French, German, Japanese, and Chinese (Simplified)" (Otter help, article updated 2026-05-06, checked 2026-10-08). Fireflies' API documentation lists Hebrew with the code "he" (Fireflies language codes, checked 2026-10-08).
2. Does it handle mixing? Many Hebrew-speaking teams switch to English for product terms mid-sentence. Some tools need you to pick one language per meeting. Test a real mixed call.

Being on the list doesn't make a tool accurate. We plan a separate Hebrew/multilingual test; results will go on the Hebrew & multilingual transcription ranking.

6. Free plans: read the limits

Free tiers are useful for trying a tool. They differ in what runs out first. Some examples, as listed on official pricing pages (checked 2026-10-08; these change often):

• Read AI Free: 5 meeting transcripts per month, meetings up to 1 hour (read.ai/plans-pricing).
• Otter Basic: 300 monthly transcription minutes (otter.ai/pricing).
• tl;dv Free: unlimited recordings and transcription, 10 meetings with AI notes (tldv.io pricing).
• Zoom's built-in assistant on the free Workplace Basic plan: limited summaries and note-taking for a few meetings per month (zoom.us/pricing).

These are the vendors' own descriptions, and we have not tested them yet. Check the current page before you decide.

A 15-minute shortlist method

1. Pick 2–3 tools that run on your platform (Zoom/Meet/Teams, your OS) and list your language.
2. Read the training and retention lines in each privacy policy.
3. Run the same recorded 10-minute call through each.
4. Check three numbers, two names and every action item against the audio.
5. Keep the one that got the important details right. A nicer summary doesn't make up for wrong details.

Related

• AI meeting note taker ranking (results planned for the second half of October 2026)
• Transcription for interviews & podcasts
• All transcription & meeting-notes tools
• How we test AI tools