Claude Meetings: An AI Teammate Joins Your Video Calls

Posted on by
Contents show
Claude Meetings

Anthropic has launched Claude Meetings in beta, and the Claude Meetings feature does something AI assistants have mostly simulated until now: Claude joins your video call as a visible participant, follows the conversation live, answers questions out loud or in chat, and starts doing the work while the meeting is still happening. Docs, whiteboards, notes, and action items are built during the call, not reconstructed from a recording afterward.

The Claude Meetings rollout covers Team and Enterprise plans, with Google Meet as the first supported platform and other conferencing tools coming soon. The design choices around permissions, external guests, and data retention reveal what Anthropic thinks an AI teammate should be: a participant that works under the same organizational rules as the people it works for, with boundaries that default closed rather than open.

This guide walks through what Claude Meetings does during a call, what it does in the meeting, what it leaves behind afterward, the security model that governs who can bring it and where it refuses to go, the pricing during beta, and where it sits in the crowded meeting-assistant market.

What Claude actually does in the call

The core loop of Claude Meetings is assistance that happens in the meeting rather than after it, and a concrete example from the launch shows the flow.

A team is planning a customer event with Claude on the call. Someone says out loud, start a doc for this plan, and Claude writes up what the team has said so far while teammates on the invite edit the document in real time. When the conversation turns to budget, a participant asks what last year's event cost; Claude asks for approval to look it up through the connected tools, posts the figure in the meeting chat, and compares it to the current plan. When timelines come up, someone asks for a whiteboard, and the team works the schedule alongside the doc. Before the call ends, a recap is requested and scheduled.

Three capabilities make the Claude Meetings loop work. First, live artifact building: docs, whiteboards, and slides are created and edited during the call, so the work product exists before the meeting ends. Second, tool-connected answers: Claude can use the connected tools of the person asking, with that person's permission, which means the answer can include real company data rather than generalities. Third, automatic documentation: notes, summaries, and action items are written continuously, removing the notetaker role from a human attendee.

Questions can be asked out loud or in the meeting chat. By default Claude answers in writing, and an answer posted in the call's chat is visible to everyone on the call, which keeps the assistance public rather than creating a private channel inside the meeting.

After the call: the Claude Meetings meeting page

Everything Claude touches during a Claude Meetings session lands on a dedicated meeting page in claude.ai, and the after-call experience is where the feature compounds.

The page holds the notes with a summary and action items, the full transcript, and every artifact built during the call. People from your organization who were on the invite receive a recap email with a link, can see the notes, and can edit what Claude made, and the page can be shared with others in the organization. The calendar view in Meetings shows past and upcoming meetings, so a team can revisit any call Claude joined or add Claude to the next one, and in a regular chat you can ask Claude questions about past meetings you have access to.

That last capability is the quiet strategic one. Once meetings become queryable documents rather than ephemeral conversations, organizational memory stops living in a rotating cast of people's notebooks and starts living in a searchable, permission-scoped archive. A new team member asking what did we decide about the event budget in October gets an answer grounded in the actual transcript, with the artifacts attached.

Anthropic is also working with meeting-tool partners including Gong, Attention, Granola, Fireflies.ai, Otter, and Plaud, so transcripts will soon flow into the revenue-intelligence and notetaking tools teams already use, positioning Claude Meetings as a source inside the existing workflow rather than a replacement for it.

The Claude Meetings security and privacy model

The boundary design is the most considered part of the Claude Meetings security model, and it defaults closed in every place a leak could happen.

External guests are the first boundary. By default, Claude stays out of calls when someone from outside the organization is invited, and it leaves if an outsider joins partway through, including another company's notetaker bot. An owner can allow Claude on calls with outside guests: on Team plans that is possible by default with an owner able to turn it off, while on Enterprise plans the owner must allow it first. On external-guest calls, Claude listens and takes notes, posts its joining notice in chat but does not reply there, and guests receive neither the meeting page nor the recap email, though what they say is included in the notes and transcript.

Data capture is the second boundary. When Claude joins, it appears as a visible participant and posts a notice in the call's chat, and it saves a transcript, a summary, and screen descriptions, not audio. The absence of raw audio recording narrows the compliance surface meaningfully, because a transcript-plus-description artifact is easier to govern than a stored recording, and participants can see exactly what is being captured.

Tool permissions are the third boundary. In Claude Meetings, Claude uses each person's connected tools with that person's permission, asks the first time, and a yes applies going forward. When someone asks out loud, Claude asks each time to confirm who is speaking, a speaker-verification step that prevents a voice in the room from spending someone else's tool access. The meeting page is visible to organization members from the invite, editable by them, and shareable within the organization, so the audience of every artifact is bounded by membership.

Transcription consent remains a human decision. Anthropic's guidance is explicit: before adding Claude or turning on auto-join, make sure everyone on your calls is OK with being transcribed. The feature is powerful precisely because it captures everything, which is exactly why the consent conversation comes first.

Admin controls and availability

The Claude Meetings deployment model differs by plan, and the admin surface determines who can bring an AI participant into a call.

On Team plans, Claude Meetings is on by default and an owner can turn it off, and inviting Claude to calls with external guests is likewise default-on with an owner off-switch. On Enterprise plans, nothing is on until an owner turns it on, either for the whole organization or for part of it through custom roles, with separate capabilities for Claude Meetings and for Meetings with external guests.

Individual control survives the admin layer: auto-join is a personal setting, letting a user add Claude to all their meetings or all except one-on-ones, and owners cannot set auto-join for another person. Claude usually joins without requiring approval, but if it asks, the host decides, and a declined join means Claude does not retry that call unless someone explicitly adds it again, with the decline visible in the interface.

The beta excludes some regulated configurations: organizations with zero data retention arrangements, HIPAA-ready configuration, or customer-managed encryption keys cannot use Claude Meetings while it is in beta, a restriction that signals which compliance surfaces the feature has not yet cleared rather than a permanent limitation.

Pricing during beta

The Claude Meetings beta economics are straightforward, and knowing what counts against usage prevents surprises.

Notes, transcripts, and answers that do not use connected tools are included free on Team and Enterprise plans during the beta. Answers that use connected tools count toward the asking person's plan usage, because those answers consume real tool operations such as document lookups. Docs, whiteboards, and other artifacts built in meetings are free through February 10, 2027, after which artifact creation presumably folds into normal plan mechanics.

For teams evaluating the feature, the practical cost model is that the documentation layer is free and the tool-using layer is metered, which aligns expense with the calls where Claude is actually doing organization-specific work rather than sitting through status updates.

Where it sits in the market

Meeting assistants are a crowded category, and Claude Meetings enters with a different thesis than the incumbents.

The notetaker bots, the category Fireflies and Otter built, record and summarize after the fact; their product is the transcript and the recap. Claude Meetings is a participant, not a recorder: it contributes during the call, builds the artifacts the meeting exists to produce, and documents as a side effect. The distinction matters commercially, because a notetaker competes on transcription accuracy while a participant competes on work done, and the second is a much larger prize.

Against the model-provider competitors, the agent-in-meeting pattern is converging from several directions, and Anthropic's differentiators are the boundary design, the organization-scoped permission model it inherits from Claude's enterprise controls, and the artifact pipeline, since docs and whiteboards built in the call are the same artifacts Claude works in everywhere.

The risks are the ones every always-on participant carries: consent management across organizations that meet each other, the accuracy of live capture in noisy multi-speaker audio, and the adoption question of whether teams actually want a visible AI in the room or prefer invisible capture. The external-guest default, leaving the call when an outsider appears, is the most conservative answer in the market to the first risk, and it will be watched as the precedent.

For readers tracking the AI-coworker pattern, the self-hosted version of the same idea is examined in our OpenBot 1.0 analysis, the real-time voice layer that makes spoken meeting interaction possible is covered in our Fish Audio Live 3 piece, and the consumer voice-assistant end of the spectrum is reviewed in our Google Home speaker review.

Getting started, step by step

Adoption of Claude Meetings is deliberately simple, and the setup path from decision to first call is short.

An owner sets the plan-level switch: Team owners confirm the default-on state or turn it off, Enterprise owners enable the feature for the whole organization or for specific roles, and both decide the external-guest policy. Individual users then set auto-join in their settings, choosing all meetings, all except one-on-ones, or manual addition only, and since owners cannot set auto-join for another person, the always-on choice stays personal.

The first Claude Meetings call follows the join flow: Claude appears as a participant, posts its notice in the call chat, and waits for direction. The launch guidance suggests starting explicit, saying Claude, start a doc for this plan, and the spoken or chat-asked instructions from there follow the natural shape of the meeting rather than a command syntax. When Claude requests tool approval, the yes binds to the person who approved, and the boundary remains that person's own access.

After the call, the meeting page is the artifact to review: checking the summary and action items against what the team actually decided is the fastest way to build trust in the documentation, and editing what Claude made trains nobody but improves the record immediately. Teams adopting the feature should run the consent conversation before the second call, because the first call demonstrates exactly what capture looks like and makes that conversation concrete.

The organizational memory angle

The longest-term consequence of Claude Meetings may be what it does to institutional knowledge, and the meeting archive is worth examining as a system rather than a convenience.

Every Claude Meetings call becomes a structured document: a transcript, a summary, action items, and the artifacts the meeting produced, all on a page scoped to the people invited. Multiplied across an organization's meeting volume, that is a queryable corpus of decisions with their reasoning attached, and the ability to ask Claude about past meetings you can access turns the archive into an interface rather than a filing cabinet.

The permission model is what makes the archive safe enough to be useful. Access follows invite membership, editing follows organization membership, sharing stays inside the organization, and the external-guest boundary keeps counterparties out of the record entirely unless an owner explicitly allows their inclusion. A decision made in a customer call is searchable by the team that made it and invisible to the customer, which is the correct default for nearly every commercial conversation.

The Claude Meetings partner integrations extend the archive's reach into systems of record: transcripts flowing to revenue-intelligence platforms mean the meeting corpus connects to the CRM, and the notetaking tools that plug into it mean teams keep their existing review workflows. The archive becomes infrastructure, and infrastructure is where the durable value of an enterprise feature accumulates.

The open question is search quality across hundreds of meetings and whether cross-meeting synthesis, asking Claude what the team has decided about pricing across the last quarter, works as well as single-meeting recall. The chat interface already supports asking about past meetings, and the capability's depth will determine whether the archive becomes a decision database or a well-organized pile.

Voice interaction in Claude Meetings and what comes next

The spoken dimension of Claude Meetings deserves its own examination, because voice is both the feature's most natural interface and its hardest engineering problem.

Participants ask questions out loud during the call, and by default Claude answers in writing, in the chat or in the artifacts it is building. The written-first default is a deliberate choice: written answers are visible to everyone, quotable in the notes, and do not clobber the human conversation the way a talking assistant would. The voice channel still matters on the input side, because saying Claude, put the timeline on the whiteboard mid-discussion is faster than typing it, and the speaker-verification step, asking each time to confirm who is speaking when someone asks out loud, is what lets the system bind a spoken request to the right person's tool permissions.

Multi-speaker attribution in real conference audio is genuinely difficult: crosstalk, accents, room echo, and overlapping speech all erode speaker identification, and the cost of a wrong attribution is a tool permission spent by the wrong identity. Asking every time is the conservative answer, and it trades a little friction for a hard guarantee.

The roadmap visible from the beta is platform expansion beyond Google Meet, transcript delivery to the partner tools teams already use, and artifact pricing folding into normal plan mechanics after February 2027. What is worth watching is whether meeting participation becomes a general capability of workplace AI, with assistants that attend, contribute, and document as a normal part of organizational life, or stays a discrete feature with its own toggle. The boundary design Anthropic shipped, defaults closed, visible participation, personal permission binding, reads like a template the rest of the industry will be compared against.

Questions readers are asking

Which plans include Claude Meetings?

Claude Team and Enterprise plans, currently in beta with Google Meet as the first supported platform. Team plans have it on by default with an owner off-switch; Enterprise owners turn it on for all or part of the organization through custom roles.

Does Claude record the audio of the meeting?

No. It saves a transcript, a summary, and screen descriptions, not audio, and it announces itself as a participant with a notice in the call chat when it joins.

What happens when an outsider joins the call?

By default Claude stays out of calls with external invitees and leaves if an outsider, including another company's notetaker bot, joins partway. An owner can allow external-guest calls, where Claude takes notes but does not reply in chat and guests get no meeting page or recap.

Is Claude Meetings free during the beta?

Notes, transcripts, and non-tool answers are included on Team and Enterprise plans. Answers using connected tools count toward the user's plan usage, and meeting-built docs and whiteboards are free through February 10, 2027.

Sources
[fluentform id="2"]