AI Notetaker vs AI Meeting Agent: What's Actually Different

An AI notetaker joins your meeting and records it. An AI meeting agent joins the meeting instead of you and runs it. The notetaker needs you in the room; the agent exists because you're not.

That's the entire distinction, and every other difference — whether it can see a shared screen, whether it pursues goals, whether it calls your systems mid-call, whether anyone gets scored — falls out of it. Once software is a participant rather than an observer, a different set of problems becomes its job.

Both categories are useful. Most teams should have a notetaker. Fewer teams need an agent, and the ones that do usually can't tell from the marketing, because everything in this space is described as an "AI meeting assistant."

The ladder

Modern notetakers are better than that table makes them sound — many do speaker separation, topic detection, CRM field updates, and searchable archives across every call your company has ever had. That's genuinely valuable and it's not what I'm arguing against.

The distinction is participation. Everything in the right-hand column requires the software to be in the conversation, making decisions while it's happening.

What a notetaker is actually for

Worth saying plainly, because posts like this usually pretend the incumbent category is useless.

A notetaker solves recall and coverage. You were in the meeting; you were also half-listening while thinking about the next one. It captures what was said, who said it, and what got committed to, and it does that across every meeting in the company without anyone changing their behaviour. Six months later you can search for the call where a customer described the workflow you're now building.

That's a real product with a real return, and it's cheap. If you don't have one, get one. Nothing below is an argument for replacing it.

The limit is structural: the meeting still consumed everyone who was in it. The notetaker made the hour more useful. It didn't give you the hour back.

What changes when the AI is a participant

The moment software joins a call as a participant rather than a recorder, it inherits problems a recorder never had.

It has to take turns. A recorder is never wrong about when to speak. A participant is wrong constantly unless someone has done serious work on it — a couple hundred milliseconds too eager and it steps on people, too slow and the silence reads as broken software. This is the single failure mode users won't forgive, and it doesn't appear on any feature comparison.

It has to hold a position. Over thirty minutes, a model with no anchoring drifts toward generic helpfulness. An agent representing you in a conversation has to still be representing you at minute twenty-eight.

It has to see. Recording a screen share is storage. Reading a screen share is a different capability: following what someone is doing in an editor, noticing they've gone quiet on a diagram, reacting to the confusion on their face. Our personas read expressions, posture, and environment through the camera, and documents, spreadsheets, application interfaces, code, and diagrams through a screen share — live, while the conversation continues.

It has to end somewhere. A recorder ends when the call ends. An agent has goals and a time budget, which means it has to decide what to cut when the conversation runs long.

Consent stops being a checkbox and becomes a design problem. When it's a recorder, disclosure is a notice. When it's a participant standing in for a person, what you say up front matters much more. We put a modal before the device check with a message you write in plain text or markdown, and the participant either accepts or leaves. One honest caveat: that click is enforced but isn't written to an auditable consent table, so if you need formal records of who agreed to what and when, you'll need your own log.

The output is the tell

The clearest way to work out which product someone is selling: ask what lands in your inbox afterwards.

A notetaker returns notes about a meeting you attended. Useful, and still an input — somebody reads it and decides what it means.

A meeting agent returns a result from a meeting you didn't attend. In our case that's the transcript, the recording, and an analysis scoring the participant against goals you defined on the scenario. Each goal has a name, a description of what good looks like, and scoring instructions; each is scored on a scale, typically 1 to 10, with written feedback, an overall average, and transcript highlights linked to the goal each supports.

That's the difference between "here's what happened" and "here's how it went, and here's the evidence." One is a document. The other is a decision you can act on without watching anything back.

The practical note for anyone writing those goals: behavioural instructions work, abstract ones don't. "Names a specific metric when asked how they'd measure success" scores consistently. "Demonstrates strategic thinking" produces noise.

Where a meeting agent actually earns its place

Not everywhere. The shapes I've seen work:

Screening conversations before a human's time is worth spending. First-round interviews, vendor qualification, inbound discovery. High volume, structured, and every one currently costs somebody thirty minutes.

Practice and assessment. Sales roleplay, onboarding checks, training scenarios. The value is the rubric — nobody has time to score twenty roleplays a week by hand.

Recurring 1:1s and check-ins you can't make. The skip-level that gets cancelled three times in a row, or the customer check-in during a launch week.

Feedback and testimonials at the moment they'd actually be given. People are far more expressive answering an AI than recording themselves cold into a webcam — we see roughly ten times the expressiveness — because the friction of self-recording removes most of what people would have said.

What these share: the meeting is structured, and its value is in the outcome rather than the relationship. Nobody should be sending an agent to a difficult conversation with a direct report. That's not a technical limitation.

Most teams want both

The honest answer for almost everyone is that these aren't competitors. They solve adjacent problems on the same calendar.

Run a notetaker across the meetings you attend, so nothing is lost and everything is searchable. Send an agent to the structured, repeatable calls that consume your week without needing your judgement. The notetaker makes your meetings more useful; the agent removes some of them.

If you have to pick one, pick by which problem is bigger. If your problem is "I can't remember what was decided," that's a notetaker. If it's "I'm in eleven meetings a day and six of them are the same meeting," a notetaker will document the problem very thoroughly and change nothing.

Frequently asked questions

Can an AI actually attend a meeting for me? Yes, and the mechanic matters. In our case the meeting happens on Waterr — you send an agent by share link, open for a window, or by scheduling it so the link is only live during the slot. The persona runs the scenario, and the transcript, analysis, and recording come back when the call ends. You can add per-send instructions — emphasise a particular skill in this one — without editing the underlying scenario.

Is an AI notetaker the same as an AI meeting assistant? "Assistant" is used for both, which is why the category is confusing. Ask one question instead: does it require you to be in the meeting? If yes, it's a notetaker regardless of what it's called.

Do participants know they're talking to an AI? Yes. There's a consent modal before the device check carrying your disclosure message — commonly used to state that the session is recorded, transcribed, and reviewed, which is required in many two-party-consent jurisdictions. In practice, hiding it would also fail immediately; people can tell.

Does the agent remember someone from a previous call? It can. Participant memory captures facts they shared — role, company, what they're working on — plus stated preferences and loose threads they didn't finish, and injects that as ambient context before the next call rather than reciting it back. Two limits worth knowing: it's off by default and enabled per scenario, and in v1 only signed-in participants are remembered, so guests joining by public link or embed aren't.

What happens if the conversation goes off-script? The agent is pursuing goals rather than reading a script, so it can follow a tangent and steer back. The time budget is what keeps that bounded — it decides what to drop when the conversation runs long.

Can I keep my notetaker and use an agent too? Yes, and most teams should. They apply to different meetings.

More on the delegation mechanic in send an agent — the better way to run 1:1 meetings, and on what this does to team coordination in the coordination tax.