# What is an AI meeting agent? > An AI meeting agent is an AI participant in a live meeting: it listens, speaks, presents, and acts, unlike a notetaker that records and summarizes. Definition, capabilities, and how one joins a call. Canonical: https://mainroom.ai/blog/what-is-an-ai-meeting-agent/ Updated: 2026-09-17 An **AI meeting agent** is an AI system that joins a live meeting as a participant and takes part in it: it listens to the room, speaks when addressed, puts material on the shared screen, looks things up, and acts on what is decided. That is the difference from an AI notetaker, which records a meeting and produces a summary afterwards. ## The short definition An AI meeting agent has four properties that a notetaker does not: 1. **It is in the room.** It has a name in the participant list, a camera tile, and a voice. People address it the way they address a colleague. 2. **It acts during the meeting, not after.** It answers a question, pulls up a slide, runs a demo, or opens a poll while the conversation is happening. 3. **It has manners.** It waits for a gap, stops when a person talks over it, goes on mute when told to, and raises a hand instead of interrupting. 4. **It has tools.** It can search the web and the team's own systems, present on the shared screen, and draft follow-ups that a human approves. The recap at the end is a by-product. Notetakers make notes; agents make the meeting itself go differently. ## What an AI meeting agent can do in a call Here is what a Mainroom agent does today, in a real Google Meet, Microsoft Teams or Zoom call, with nothing installed on the platform: - **Speak and listen.** Joins as a participant with a camera tile and a voice. Answers when addressed, waits for a gap, and never talks over a person. - **Slides and pages on the shared screen.** Puts up text slides, comparison tables, charts, timelines and small calculators on the meeting's shared screen, and takes them down when asked. - **Whiteboard.** Teaches on a shared whiteboard it draws on step by step: boxes, arrows, notes, highlights. Questions about the drawing are answered on the drawing. - **Live browser demos.** Opens a real browser on the shared screen, follows instructions from the room ("click the pricing tab", "scroll down"), and narrates. - **Lookups in connected tools and the web.** Runs a background search in the team's connected tools (any MCP server) or the web, and reports when there is a gap in conversation. - **Generated images.** Generates an image from a description and puts it on the shared screen. - **Anonymous intake.** Opens an anonymous response link for the room, shows a live count, and synthesizes the responses on request. - **Meeting chat and private notes.** Posts links and recaps to the meeting chat; sends a private note to one person on request. - **Meeting ledger.** Logs decisions, commitments with owners and dates, open questions and risks the moment they are confirmed, and carries them to the next meeting with the same people. - **Proposed actions.** Drafts CRM notes, tickets and follow-ups as proposals the owner approves before anything is written to a system of record. - **Live coaching.** Sends private nudges and answers private questions on a companion page only the inviter can see. - **Memory.** Remembers people and prior sessions across calls, per agent, as a switch the owner controls. - **Post-call email.** Emails a role-appropriate recap after the call: a coaching debrief, an account summary, or a team recap. - **Floor control by voice.** Goes on mute when told to and stays silent until asked to come off mute. Raises a hand instead of interrupting. ## How an AI meeting agent joins a meeting On Mainroom the agent has its own email address. You add that address to a calendar invite, or forward the invite to it, and the agent joins at the start time. It appears in the lobby and the host admits it, exactly like a guest. Captions on the platform tell it who is speaking by name. Under the hood the agent runs on a voice model that listens and speaks in real time (OpenAI's realtime voice models, in our case), with a gate in front of the model that decides when a reply is wanted at all. That gate is most of the work. A model that answers everything it hears is unusable in a room of people talking to each other, so the agent tracks who was addressed, who is in the meeting, whether it is a one-on-one or a room, and whether it has been told to be quiet. ## When you would want one The honest list, from our own use: - **Customer calls**, where an agent remembers the account, logs commitments both ways, and coaches the account owner privately without the customer seeing anything. - **Recurring team meetings**, where the agent confirms decisions, asks for owners and dates as actions land, and brings back last week's open items. - **Teaching and explaining**, where the agent draws on a whiteboard step by step instead of describing a diagram in the air. - **Demos**, where the agent drives a real browser on the shared screen and follows instructions from the room. - **Getting the quiet half of the room heard**, with an anonymous intake link and a synthesis of the responses. ## When you would not An agent is a participant, and participants have a cost: attention. If a meeting is a private conversation between two people who know each other well, a notetaker that stays out of the way may be the better tool. Mainroom's Silent Coach template exists for exactly that case: it joins as a plain notetaker, never speaks, and coaches the person who invited it privately. ## What to look for when evaluating one - Does it join the platforms you use, as a participant, without a plugin? - Can you tell it to be quiet, and does it stay quiet? - Does it act on the shared screen, or only in a side panel? - What can it connect to, and does anything get written to your systems without approval? - Who sees the private coaching, if there is any? - Can you build your own agent, with its own persona, tools and manners, or only use a fixed assistant? Mainroom is one answer to these questions. The [comparison pages](/compare/) go through the alternatives. ## Frequently asked questions **Is an AI meeting agent the same as an AI notetaker?** No. A notetaker records the meeting and produces notes afterwards. A meeting agent takes part in the meeting while it is happening: it answers, presents, looks things up, and gathers input. Most agents also produce notes, but that is not what makes them agents. **Does an AI meeting agent need a plugin or app installed?** Not on Mainroom. The agent joins Google Meet, Microsoft Teams or Zoom as a regular participant from a calendar invite, and the host admits it from the lobby. **Can everyone in the meeting see the agent?** Yes. It appears in the participant list with its own name and camera tile, and the platform shows its standard notice. It never joins silently.