Due to increased demand, text TeamShift to hold the next available slot.+1 717 740 8200Call instead
TeamShift

Guide

AI notetakers: how they work and what to check before you hit record

Direct answer

An AI notetaker records a meeting, converts the speech to text, and uses a language model to write a summary, action items, and searchable notes. It either joins the call as a bot, listens to your computer's audio, or runs inside Zoom, Teams, or Meet. Before you use one, tell everyone they're being recorded (about a dozen states, including California and Pennsylvania, require every party's consent), check the summary against what was actually said, and read how the vendor stores your recordings. If you're picking a product, our AI meeting assistant roundup compares prices.

How an AI notetaker works

Every notetaker runs the same basic pipeline:

  1. Capture. The tool gets the audio, either through a bot that joins the call, from your device's microphone and speakers, or from the meeting platform itself.
  2. Transcribe. Speech recognition turns audio into text. Most tools label speakers ("Speaker 1," or the person's name if it can match their voice or their meeting login).
  3. Summarize. A language model reads the transcript and writes the summary, decisions, and action items. Many tools let you pick a template, such as a sales call or a one-on-one.
  4. Store and share. The recording, transcript, and summary are saved in the vendor's cloud. Many tools email the recap to attendees or push it to a CRM or Slack.

The "AI meeting notes" you see in your inbox after a call come out of step 3. They're only as good as the transcript underneath them.

Bot vs. botless notetakers

Bot joins the call Botless (device audio) Built into the platform
Examples Otter, Fireflies, Read AI, tl;dv Granola, Fathom's bot-free option (beta) Zoom, Teams Copilot, Meet with Gemini
Others see it Yes, as a named participant No, unless you say so Usually a recording or notes banner
Works on any platform Most major ones Anything your computer can hear Only that platform
Can be blocked by host Yes, hosts can refuse entry No Controlled by the account admin
Consent burden Bot gives some notice, not full consent Entirely on you Platform notice helps

A visible bot is a useful signal, but it isn't consent. People may not notice it, may not know what it does, or may not realize the vendor keeps the audio. Botless tools are less awkward in client meetings, and that's exactly why you have to say out loud that you're recording.

Watch auto-join settings. Many bot-based tools connect to your calendar and join meetings on their own. Fireflies, for example, lets you choose between joining every meeting with a conference link, only meetings you own, only meetings with teammates, or only when you invite it. If you set "all meetings," the bot can turn up in calls you forgot about, including other people's.

Recording consent law

This is general information, not legal advice. Talk to a lawyer if you record calls with people in other states or countries.

Federal law allows recording when one party to the conversation consents. The federal wiretap law, 18 U.S.C. 2511(2)(d), says it's not unlawful for a party to a communication, or someone with one party's prior consent, to intercept it.

Some states require everyone's consent. California's Penal Code 632 bars recording a confidential communication "without the consent of all parties," with fines up to $2,500 per violation for a first offense. Pennsylvania, Florida, Illinois, Maryland, Massachusetts, Montana, New Hampshire, and Washington are also commonly listed as all-party consent states. Lists differ at the edges (Connecticut, Michigan, Nevada, Oregon, and Delaware come up depending on the type of conversation), so treat any count as approximate. When people on the call are in different states, the safe rule is to follow the strictest one.

Voiceprints are a separate issue. Illinois' Biometric Information Privacy Act treats voiceprints as biometric identifiers. Two class actions have been filed against Fireflies under that law, including Fricker v. Fireflies.AI Corp. in March 2026, alleging the tool collected voiceprints of participants who never had accounts. Otter faces a class action, Brewer v. Otter.ai, filed in August 2025, alleging its notetaker recorded non-users without consent and used the recordings for model training. Jackson Lewis summarized the complaint. These are allegations, not findings, but they show where the risk sits: with the people on the call who never signed up.

What to do in practice:

  • Say at the start: "I'm using an AI notetaker to record and transcribe this. Is everyone OK with that?" Stop if anyone objects.
  • Put a line in the calendar invite saying the meeting will be recorded and transcribed.
  • Turn off auto-join for external meetings, or limit it to meetings you host.
  • Don't record HR, legal, medical, or disciplinary conversations unless your counsel says it's fine.

Accuracy limits

Transcripts are good enough to search and skim, not good enough to quote without checking. The weak spots are predictable:

  • Names and jargon. Product names, part numbers, and people's names get mangled. Several tools let you add a custom vocabulary, and Otter's Pro plan lists team vocabulary.
  • Crosstalk and bad audio. When two people talk at once or someone is on a speakerphone in a truck, words drop out.
  • Speaker labels. Tools mix up who said what, especially when people sound alike or share a room. Speaker attribution and action items assigned to the wrong person are recurring complaints in Fireflies' Trustpilot reviews.
  • Summaries that sound sure. The summary may state a decision that was only floated, or skip a caveat. Check numbers, dates, and commitments against the transcript before acting on them.
  • Length and language caps. Otter's free plan stops at 30 minutes per conversation and covers six languages. Fireflies caps recording length at two hours on its Free and Pro plans, per its pricing page.

A simple habit fixes most of this: before the recap goes to a client, the person who ran the meeting reads it and corrects it.

Privacy and data retention

Your recordings sit on the vendor's servers, so their terms are your terms. Check four things:

  1. Training. Does the vendor train AI on your recordings? Otter's privacy policy says it uses de-identified audio and transcripts to train its models. Fireflies' pricing page says your data is never used for AI training. Granola lets each user opt out of model training, and its Enterprise plan adds a workspace-wide opt-out.
  2. Retention. How long are recordings kept, and can you set a limit? Custom retention is usually an enterprise feature: Fireflies Enterprise, Read AI Enterprise+, and Fathom Enterprise all list it.
  3. Sharing. Who gets the recap by default? Some tools email every attendee. Make sure internal notes don't go to a client.
  4. Regulated data. If a meeting might include patient information, you need a plan that signs a Business Associate Agreement. Otter (Enterprise add-on), Fireflies (Enterprise), Read AI (Enterprise+), and Granola (Enterprise) list HIPAA options. Free and low tiers don't.

Also delete what you don't need. A transcript you no longer keep can't leak or be subpoenaed.

Using an AI note taker for Zoom

For Zoom specifically you have two routes. Zoom's own AI features run inside the meeting, and its free Workplace Basic plan includes meeting summaries for three hosted meetings a month. Or you invite a third-party bot (Otter, Fireflies, Fathom, Read AI) or run a botless app like Granola. Zoom hosts can enable a waiting room and refuse bots they don't recognize, so if you're the guest, ask the host before your bot shows up.

After the meeting

Notes are only useful if someone acts on them. TeamShift isn't a notetaker: it doesn't record meetings, keep audio or video, or produce transcripts. It can join a Zoom, Meet, or Teams call as a clearly disclosed AI participant, shown with an avatar or a labeled tile, and speak for your business within the brief you give it. The follow-up is drafted for your OK. The Sales worker drafts the recap email or proposal, and the Operations lead turns next steps into tasks and checks they got done. Nothing goes to a client until you approve it. See human-in-the-loop AI for how that approval step works.

FAQ

Is it legal to use an AI notetaker without telling people?

Not everywhere. Federal law and most states allow recording with one party's consent, but California, Pennsylvania, and other all-party consent states require everyone's agreement. The safe practice is to announce the recording at the start of every call.

How accurate are AI meeting notes?

Transcripts are usually good enough to search and skim, but they miss names, jargon, and crosstalk, and they mix up speakers. Summaries can state things more firmly than people said them. Read the recap before you send it or act on it.

What's the difference between a bot and a botless notetaker?

A bot joins the call as a visible participant and records from inside the meeting. A botless tool, like Granola, records your computer's audio so no one else sees it. Botless is less awkward but puts the whole job of getting consent on you.

Do AI notetakers train on my meetings?

It depends on the vendor. Otter's privacy policy says it trains on de-identified recordings and transcripts. Fireflies says it doesn't use your data for training. Read the privacy policy and look for an opt-out before you record client calls.

Can I use an AI notetaker for medical or legal meetings?

Only with care. For patient information you need a plan that signs a Business Associate Agreement, which is typically an enterprise tier. For legal or HR conversations, ask counsel first, since a transcript becomes a record that can be requested later.