Of AI Meeting Guests and the Calendars They Can’t See

Posted by & filed under , .

Illustration of three people beside a giant wall calendar, one crossing off days with an oversized pencil.So Anthropic announced Claude Meetings on 12 November: a beta for Team and Enterprise plans where Claude joins your Google Meet call as a visible participant, takes notes, answers questions in the chat and even builds documents and whiteboards while people are still talking. You can add it to an upcoming meeting straight from your calendar, or switch on auto-join. I can already hear half of LinkedIn saying “finally, someone else can go to my meetings” 🙂

Tempting, I’ll give you that. But I think the interesting part of this announcement isn’t the note taking (we’ve had bots doing transcripts for years, mostly sitting in calls nobody remembers inviting them to). It’s the fact that the agent now gets into your meetings through your calendar. And that’s where I, as someone building a calendar product, start squinting a bit.

The Agent Is Only as Good as the Calendar It Reads

Here’s the thing: most people I talk to nowadays don’t have one calendar. They have a work Google Workspace account, a Microsoft 365 account a client gave them as a contractor, a personal Gmail, and quite likely a side-project calendar they forgot they created. An agent that joins meetings and then (inevitably, because this is where it’s all going) books the follow-ups sees exactly one of those. So it happily proposes “Thursday at 2pm” for the follow-up, while you’re on a call with another client in a calendar it has never heard of. Double booking, now with added AI!

This is exactly the problem that got me building Calendrz: availability mirroring across connected accounts, so every calendar shows you as busy when any of them does. And when we built the Calendrz smart scheduler (an agent that books meetings from a chat prompt or a forwarded email thread), the first limitation we wrote down in our own docs was precisely this one: the agent only knows when someone is free if it can see their calendar. Anything else is the agent guessing, and guessing confidently, which is worse.

Meetings Are Becoming an API

The bigger shift, I’d argue, is that the meeting itself is turning into something agents read and write. Claude asks before using each person’s connected tools for the first time, it posts its answers where everyone on the call can see them, and by default it stays out of calls with outside guests. Sensible defaults, by the way: I suspect somebody had to sit through a demo where the bot cheerfully summarised the pricing strategy for the vendor on the call.

But it also tells you where things are heading. I genuinely believe the UI is on its way out: your agent does the clicking, filling in and navigating for you, and the calendar grid you stare at every morning is one of those UIs. In a couple of years I suspect most meetings will be proposed, booked, moved and summarised by agents, and humans will mostly just turn up (or not). Which is why in Calendrz we built an MCP server early on, so Claude or ChatGPT can read your events and connected accounts directly, rather than squinting at a screen the way I do.

What I’d Do as a CTO on Monday

If you’re rolling something like this out in your company, a few practical bits:

  • Decide the consent rules before the first auto-join, not after. Transcripts of every meeting are a lot of data to have lying around, and “who said Claude could be here?” is not a conversation you want to have in front of a client.
  • Treat the agent’s access as a permissions problem, not a feature toggle. Which calendars can it see? Which tools can it call, and on whose behalf?
  • Ask the obvious question: if an agent can attend this meeting and produce its output, does the meeting need to exist at all? Some of them, truth be told, should have been a document.

Bottom line: an AI that joins your meetings is useful. An AI that books your next meeting while seeing only a third of your week is a very polite way of creating chaos. Fix the calendar first, then invite the robot 😉