No Bot, No Saved Audio, One-Click Follow-Ups in ChatGPT Meetings

What OpenAI shipped

OpenAI has added a Meetings plugin to the ChatGPT desktop app for Mac. According to the OpenAI Help Center, it takes notes during online calls or in-person conversations and saves a personalised summary with action items in ChatGPT Space. No bot joins the call, so nothing new shows up in the participant list.

It is a separate product from ChatGPT's older Record feature. Cult of Mac explains that Meetings combines the transcript with your own notes and context from connected apps, and positions it as a direct competitor to Granola, Wispr's Notetaker and other Mac note-takers. Neowin adds that ChatGPT can detect a call in progress and offer to take notes even when the meeting is not on your calendar.

OpenAI labels it a beta, and that is the right lens. It is a first version built into a general-purpose assistant. Teams comparing it with dedicated meeting software should expect fewer controls and less documentation for now.

Who can use it today

Access depends on your device and your plan. On Business plans there is also an admin step: someone has to enable both ChatGPT Space and the Meetings plugin for the workspace before anyone can use it.

Because capture runs only in the Mac app for now, meetings attended only by Windows or mobile users will not be captured. If your team plans to search past meetings, check how much of your calendar the plugin would actually cover.

• Platform: ChatGPT desktop app on macOS, with Windows, iOS and Android listed as coming soon, per OpenAI.
• Plans: beta on all Pro and all Business plans, per OpenAI. Nerdschalk reports Enterprise access is a limited alpha arranged through account managers, and Neowin says select Enterprise customers will get access.
• Not listed: Plus, Go and Free plans, according to Nerdschalk, which also says the rollout is gradual, so eligible accounts may not see the plugin yet.
• Price and date: third-party explainer CellCog cites Pro from $100 a month and Business at $25 per user a month ($20 billed annually), and dates the release-notes entry to September 29, 2026 (DevDay). OpenAI's help page does not state these, so confirm them against your own account.

How the follow-up automation works

Based on the OpenAI help page, the flow from conversation to completed task has four steps. Each one is a place where errors can enter or be caught.

Extras around that flow: connecting Google Calendar or Outlook Calendar is optional and adds reminders and a view of upcoming meetings. You can also ask questions about past meetings, such as when you agreed to schedule a follow-up with a named colleague.

• Capture: the plugin streams microphone and system audio from your Mac to OpenAI while you take your own notes alongside it.
• Merge: after the meeting, it combines the transcript, your notes and context from connected apps you have authorised into a summary with suggested action items.
• Review: on the meeting page, you read and edit each suggested action item.
• Start: one click hands an item to ChatGPT Work, which can draft a follow-up email or update a project plan.

Treat every suggested action as a draft

The review step before anything runs is a sensible design. A wrong draft email costs you a minute to delete. A project plan updated with the wrong owner costs more, because colleagues may act on it before anyone spots the error.

I found no hands-on reports in the sources that measure how often the plugin assigns the wrong owner or records a decision nobody made. Nobody has published data on this yet, and it is the most important thing to test before relying on the automation.

A practical rule for the first weeks: only start an action item you personally heard agreed. Any summary can make a loose discussion read like a firm commitment, and the one-click button removes the pause where people used to catch that.

Where the audio goes and who sees the notes

OpenAI says the plugin streams audio to its servers to generate notes. Once the notes are ready, the audio is deleted from the Mac and from OpenAI's servers and cannot be replayed. If you are offline or processing is incomplete, audio may stay on the Mac temporarily.

Deleting the audio reduces what can leak, but it also means you keep a generated summary and no recording. If a client disputes what was promised, or an auditor asks how a board decision was reached, you cannot go back to the source. For legal, compliance and governance meetings, that is a real limitation.

On sharing, the ChatGPT Space page says Meetings notes are private by default and you choose what to share. For Business, Enterprise and Edu, OpenAI's plugins page says data from connected apps is not used for training by default, and admins control which plugins and connected apps are available. The sources say nothing about data residency, meaning where the audio is processed, so ask OpenAI directly if your policy depends on it.

Non-English and multilingual meetings

OpenAI has not published language accuracy figures for Meetings in the material reviewed. The closest reference is the help page for the older Record feature, which says it works best in English and that accuracy in other languages is improving but can vary. Whether Meetings behaves the same way is not documented.

For teams that meet in Turkish, German, Arabic or a mix of languages, the useful test is concrete: do names, figures and deadlines come through correctly? Those are the details that end up in action items. A misheard date or a mangled client name produces a follow-up that looks confident and is wrong.

Speaker attribution needs the same scrutiny. The Record help page says that feature separates speakers with generic labels like Speaker 1 that you can rename. The Meetings page does not explain how overlapping speech is handled. Since ownership of a task depends on who said what, test it on a busy call before trusting the owners it assigns.

A review by competitor Sally describes Record mode as macOS-only, primarily English, without calendar auto-join or CRM integration, and with a recording limit of about 120 minutes. It comes from a rival vendor and covers the older Record feature, not Meetings, so use it as a list of things to verify rather than a verdict.

Consent is left to the person pressing start

According to CellCog, the in-app reminder is visible only to the user and does not notify other participants or collect their consent. OpenAI's Record help page tells users to get the right consents before recording others. Either way, responsibility sits with whoever starts the capture.

Recording-consent and data-protection rules differ by jurisdiction, and calls with participants in several countries may fall under more than one set. Check with counsel before rolling the plugin out to a whole team.

With a visible bot, everyone on the call at least knows something is listening. With quiet capture on one laptop, the team needs a written habit instead: announce it at the start, note it in calendar invites, offer an opt-out, and agree in advance which external calls are never captured.

A decision checklist

The plugin works well as a personal assistant that speeds up follow-up work. It is weaker as a system of record. Use the checklist below, then confirm the answer with a two-week pilot on real calls, comparing suggested action items with notes a person took in the same meetings.

If several items in the second list apply, a dedicated meeting tool, or a self-hosted one where data must stay on your own infrastructure, is likely the better fit. Mixing is also reasonable: ChatGPT for everyday internal calls, a specialised tool for meetings that need minutes.

• ChatGPT is likely enough if: attendees use Macs, meetings are mostly in English, notes serve personal follow-up, and nobody needs replayable audio.
• Look elsewhere if: you need Windows or mobile capture today, replayable source recordings, a stated data-residency option, or formal minutes with verified speakers and owners.
• Look elsewhere if: limits on meeting length or CRM integration matter to you (see caveats above).
• Pilot before deciding if: meetings are multilingual, in-person with many speakers, or involve external participants where consent is unclear.