Do AI notetakers need a bot to join the meeting?

No. A joining bot is the most common capture method, not a requirement. A bot works by signing into the meeting as a separate participant — which is why it shows up in the attendee list — and it's popular because it works identically on Zoom, Teams, and Google Meet without installing anything on your computer. The alternative is to capture the audio your computer is already playing, using the operating system's own audio APIs. That needs a desktop app and only covers calls taken on that machine, but it adds no extra participant to the meeting. Canary uses this second approach.

Do AI notetakers transcribe in real time or after the meeting?

Almost all of them transcribe in real time — streaming speech recognition returns text within a second or two of speech, which is why you can often watch a live transcript scroll. The summary is the part that usually waits. Most notetakers buffer the whole transcript and run one summarization pass after you hang up, so the notes land in your inbox minutes later. A smaller group summarizes continuously during the call, re-summarizing the transcript as it grows so there's always a current summary to read mid-meeting.

How accurate are AI meeting notes?

Transcription accuracy on clear audio from a good microphone is high enough that most notes read cleanly, but it degrades predictably: cross-talk, heavy background noise, strong accents, and unfamiliar jargon (product names, acronyms, people's names) are where errors cluster. Summary quality depends on the transcript underneath it, so a bad-audio meeting produces a confidently wrong summary rather than an obviously broken one. Treat AI notes as a reliable memory aid and a starting point, not as a verbatim legal record — and check the transcript before quoting anyone.

How do AI meeting notetakers work? The pipeline, explained

Q: How do AI meeting notetakers work?

Every AI meeting notetaker runs the same four-stage pipeline: capture the audio, transcribe it to text with speech recognition, label who said what (speaker diarization), then feed the transcript to a language model that writes the summary and pulls out action items. The real differences between tools are in stage one and stage four. Stage one is how the audio is captured — a bot that joins the call as a participant, a platform's built-in recording, a browser extension reading live captions, or direct capture of your computer's own system audio. Stage four is when the summary is written — after the meeting ends for most tools, or continuously during the call for real-time summarizers like Canary.

Underneath the branding, every AI meeting notetaker on the market runs the same four-stage pipeline: capture the audio → transcribe it → figure out who said what → summarize it with a language model. Once you can see those four stages, the whole category becomes legible, and the marketing differences between tools resolve into two real engineering choices: how the audio gets captured, and when the summary gets written.

Stage 1: Capturing the audio

This is the stage that determines what the tool feels like to use, and there are four ways to do it.

A bot joins the call. The tool signs a separate account into your meeting — a headless client driving Zoom, Teams, or Google Meet like a participant would — and records the meeting from the inside. This is how Otter, Fireflies, Fathom, and most of the category work. It’s popular because it’s platform-agnostic and needs nothing installed on your computer; the meeting bot does all the work in the vendor’s cloud. The cost is social: the bot appears in the participant list with a name like “Fireflies Notetaker,” and it’s an extra identity holding a copy of your conversation. That’s why a bot shows up in meetings you didn’t expect it in.

The platform records it natively. Zoom AI Companion, Microsoft 365 Copilot, and Google Meet’s Gemini notes skip the bot entirely because they are the meeting platform — they tap the media server directly. Excellent fidelity, zero setup, but it only works inside that one platform, and it’s typically gated behind licensing and admin settings your IT team controls.

A browser extension reads the live captions. Tools like Tactiq don’t capture audio at all. They attach to the meeting tab and scrape the caption text the platform is already generating. It’s lightweight and bot-free, but it inherits the platform’s captions — so it only works in a browser tab, only when captions are on, and only on supported platforms.

The app captures your computer’s system audio. A desktop app asks the operating system for the audio your machine is already playing and reads it directly — CoreAudio taps on macOS, WASAPI loopback on Windows. This is system audio capture, and it’s how Canary works. Because it reads the sound at the OS level, it’s indifferent to which app the call is in — Zoom, Meet, Teams, Discord, a webinar, a recorded video — and it adds no participant to the call. Older tools did this by making you install a virtual audio device like BlackHole or Loopback; modern OS APIs make that unnecessary. The trade-off is real: it only covers meetings taken on that computer, so there’s nothing to capture calls you didn’t attend.

Stage 2: Speech to text

The captured audio is chunked and sent to a speech recognition model, which returns text. Nearly every modern notetaker uses streaming transcription: audio flows continuously and text comes back within a second or two, rather than waiting for the meeting to end and processing one big file.

Streaming models emit interim results first — a fast, provisional guess that visibly rewrites itself as more context arrives — then finalize each phrase. That flickering-and-settling behavior in a live transcript isn’t a bug; it’s the model revising its hypothesis. Most tools also run voice activity detection ahead of this to skip silence, which cuts both cost and latency.

Stage 3: Who said what

Raw transcription produces an undifferentiated wall of text. Speaker diarization segments it by voice, so the transcript reads as a conversation instead of a monologue. Tools that join the call as a participant often get a shortcut here: the platform tells them which participant’s audio stream is active, so attribution comes with a real name attached. Tools capturing a single mixed audio stream have to separate voices acoustically and end up with “Speaker 1 / Speaker 2” until you label them.

This is a genuine trade-off worth knowing about before you pick a tool. It also matters less than you’d expect for the in-meeting use case — when you’re catching up on what was just said, what was decided usually matters more than which voice said it.

Stage 4: Turning the transcript into notes

The transcript goes to a language model with instructions to produce a summary, key decisions, and action items. Most tools then do post-processing — matching the notes to a calendar event, applying a template, and pushing the result to email, Slack, Notion, or a CRM.

This stage is where the category actually splits. Two designs exist:

Batch, after the call. The tool waits for the meeting to end, then runs one summarization pass over the complete transcript. Every result is written with full knowledge of how the meeting turned out, which makes the artifact clean and well-organized. It also means nothing exists until you hang up.
Rolling, during the call. The tool re-summarizes continuously as the transcript grows, maintaining a rolling summary that’s always current. This is harder — it has to re-condense repeatedly under latency pressure, without full knowledge of where the conversation is heading — and it’s why most tools don’t do it.

Canary is built on the second design, and takes it one step further with a multi-resolution summary: the same conversation maintained simultaneously at several zoom levels — what’s being said right now, the last 2 minutes, the last 5, and the whole call. When your name is called after you’ve drifted, you read the 2-minute view and you’re back. That’s a different job from the after-the-call artifact, which is why the two designs coexist rather than one replacing the other.

The two questions that classify any notetaker

You can place any tool in the category by asking:

How does it get the audio? Joining bot, native platform, caption-scraping extension, or local system audio. This determines whether an extra participant appears in your call, which platforms it covers, and how much of the conversation leaves your machine.
When does it write the summary? After the meeting, or continuously during it. This determines whether the tool can help you in the meeting or only after it.

Most of the market is bot + after-the-meeting. Granola and Jamie are bot-free + after-the-meeting. Canary is bot-free + during-the-meeting, which is the corner the category has mostly left empty. See Canary vs Otter for the bot-based contrast and Canary vs Granola for the bot-free-but-post-meeting one, or real-time vs after-the-meeting AI notes for which timing fits your work.

What this pipeline means for privacy

Knowing the stages tells you where your meeting actually goes. Audio leaves your machine at stage 2 in almost every case — live transcription at usable quality means cloud speech recognition — and the transcript text leaves again at stage 4 for summarization. Essentially no production notetaker is fully offline today, and any tool claiming otherwise deserves scrutiny.

What varies is everything around that: whether a bot holds its own copy of the meeting, whether raw audio is stored or discarded after transcription, who the speech-to-text and LLM subprocessors are, and whether capture is always-on or starts only when you press a button. Are AI meeting notetakers safe? walks through how to check each of those in about five minutes.

And regardless of which pipeline your tool uses: tell people you’re capturing notes. No capture method removes that obligation — bot-free means a smaller footprint, not a quieter one — and consent rules for recording vary by region. Is it legal to record a meeting? covers the one-party vs two-party distinction.

Bottom line

AI meeting notetakers all capture, transcribe, diarize, and summarize. The interesting variation is at the ends of the pipeline: how the audio is captured — bot, platform, captions, or your computer’s own system audio — and whether the summary is written after the call or maintained live during it. Pick based on those two answers, not the feature list in the middle, because the middle is roughly the same everywhere.

How do AI meeting notetakers work?