AI Meeting Assistant Now Acts Mid-Call: What Changes

You finish a 45-minute planning call, and the hard part starts. Someone has to write up what was decided, who owns what, and which dates moved. Now an AI meeting assistant can sit in the call as a visible participant and build that plan while people are still talking.

That is a real shift. For years, AI in meetings meant a bot that listened, recorded and sent a summary afterward. The newest tools act during the call: they draft documents, sketch timelines and answer questions in the chat. That changes how much time you get back, and it also changes who is responsible when the bot hears something it should not have.

Here is what the new generation actually does, where the consent rules still bite, and a simple way to work out whether one pays for itself on your team.

What Claude Meetings Does During a Call

According to TechCrunch's report on the launch, Anthropic announced Claude Meetings on Tuesday, November 10. Claude joins a video call, follows the conversation and works on tasks in real time. It can draft the plan being discussed, map a timeline on a shared whiteboard, answer questions in the meeting chat and pull figures from connected work apps.

When the call ends, the notes, action items and anything Claude built are waiting in claude.ai. The launch also added two companion tools:

• Claude Whiteboards: a shared board that turns a spoken plan into timelines and deadlines, and can be converted into a document or slide deck.
• Claude Sheets: spreadsheets with tabs, formulas and dropdowns, where Claude edits cells alongside the team and explains each change.

The limits matter as much as the features. At launch it works with Google Meet only, and only on Team and Enterprise plans. All three tools are labelled beta. If your company runs on Zoom or Microsoft Teams, this one is not for you yet.

Anthropic is not alone. TechCrunch notes that OpenAI shipped workplace tools a few weeks earlier for visualizing company data and building presentations, and Meta's Muse agent builds small apps on the fly. Microsoft and Google, which own the office software most people already use, are the incumbents all three are chasing.

Why an AI Meeting Assistant That Acts Is Different

The older model was listen now, summarize later. A notetaker joined, transcribed every word and emailed a recap. Useful, but the recap still needed a human to turn it into a plan, a ticket or a budget line.

An assistant that acts mid-call collapses that gap. If someone says "let's push launch two weeks and add a QA pass," the timeline can update on screen while the group is still there to object. Mistakes get caught by the six people in the room instead of by one tired person writing notes at 6 pm.

Timing is the other reason this matters. Microsoft's Work Trend Index research on the "infinite workday" found that 57% of meetings are ad hoc calls without a calendar invite, and meetings after 8 pm rose 16% year over year. Unplanned calls rarely have a note-taker assigned, so they are exactly the ones where decisions get lost.

Where the acting model can go wrong

An assistant that writes things down can mishear. One that edits a shared spreadsheet can mishear and then change a number everyone trusts. That is a different kind of risk, and the question of who is responsible when an AI agent makes a mistake gets sharper when the agent is editing live team files.

Claude Sheets narrates each change it makes, which helps. Still, treat any figure the assistant enters during a call as a draft until a person who owns that number has checked it.

The Consent Problem Meeting Bots Created

Recording laws were written for phone calls, and they did not anticipate a bot quietly joining every meeting on a calendar. That gap has already produced a lawsuit.

The Register reported that plaintiff Justin Brewer filed a complaint against Otter.ai seeking class action status, arguing it belongs in the US District Court for the Northern District of California. The complaint says Otter Notetaker records every participant in Zoom, Google Meet and Teams calls, including people without Otter accounts, and never asks those guests for consent. It also cites Otter's own policy that participants' voices are used to train its speech recognition. The filing calls Otter's stance, that account holders should warn their guests, an attempt at "outsourcing its legal obligations to its accountholders."

Those are allegations, not findings, and the case had not been decided in anything we reviewed. But it shows the exposure.

Federal law allows recording with one party's consent, but roughly a dozen states, including California, Florida, Pennsylvania and Washington, require everyone on the call to agree. A meeting with people in three states can fall under the strictest of those rules.

How the new design answers it

Claude Meetings takes a narrower approach. It identifies itself as a participant, posts a notice in the meeting chat, refuses to join calls with external guests unless an admin authorizes it, and leaves on its own if an outside guest joins. That keeps it inside one company's walls by default, which is where consent is easiest to manage through employment policy.

It does not settle every question. TechCrunch's report does not describe how long call data is kept, so check your admin settings and the vendor's data terms before rollout. Our guide on what happens to your data when an AI app shuts down covers the questions to ask any AI vendor holding your conversations.

Doing the Math: What a Mid-Call Assistant Saves

Vendors rarely publish time-saved figures you can audit, so run your own. Here is a worked example with stated assumptions you can swap for your team's numbers.

Take a six-person team with one 45-minute planning call a week. Assume each person's loaded cost is $60 an hour.

• The meeting itself: 6 people x 0.75 hours x $60 = $270 a week.
• Follow-up today: one person spends 40 minutes writing notes and the plan. The other five each spend 10 minutes finding their action items. That is 90 minutes, or $90 a week.
• Follow-up with an assistant: if the plan is built during the call and only needs a 15-minute review plus 3 minutes per person, follow-up drops to about 30 minutes, or $30 a week.

The saving is $60 a week, or $3,120 over 52 weeks. Spread across six people, that is $520 per seat per year, or about $43 per seat per month. If the tool costs less than that and the team would use the plan anyway, it pays for itself on this one meeting alone.

Now stress-test it. If the assistant makes one error a month that takes someone an hour to unwind, subtract $720 a year. If your team only meets every other week, halve the saving. The break-even moves fast, which is why a two-week trial with a stopwatch beats any vendor slide.

How to Use One Without Getting Burned

Whichever tool you pick, a few habits keep the time savings and drop the risk.

1. Announce it out loud. A chat notice is easy to miss. Say "Claude is taking notes and drafting the plan" at the start, and offer to switch it off.
2. Keep it off external calls by default. Clients, candidates and vendors did not sign your employment policy. Get explicit agreement first.
3. Treat live edits as drafts. Budgets, dates and headcount numbers need a named owner to confirm them after the call.
4. Read the retention terms. Know whether recordings or transcripts are stored, for how long, and whether they train models.
5. Watch for scope creep. An assistant connected to your CRM or finance tools can surface that data in a room that should not see it. Limit connectors to what the meeting needs.

These tools run on a large language model, which predicts likely text rather than storing a perfect record. That makes them excellent at drafting and occasionally confident about something nobody said. Build your process around the first strength and guard against the second.

Frequently Asked Questions

Is it legal to use an AI meeting assistant to record calls?

It depends on where the participants are. Federal law permits recording with one party's consent, but about a dozen states, including California and Florida, require everyone's agreement. The safest habit is to announce the assistant and get agreement from everyone on the call.

Does Claude Meetings work with Zoom or Microsoft Teams?

Not at launch. Claude Meetings works with Google Meet only and is available on Team and Enterprise plans as a beta. TechCrunch reports that support for more plans is expected later.

What happens if an outside guest joins a Claude meeting?

Claude leaves the call automatically. It also will not join meetings with external guests in the first place unless a company admin has authorized it.

Can AI notetakers use my voice to train their models?

Some do. The lawsuit against Otter.ai cites Otter's own policy that it uses participants' voices to train its speech recognition. Check each vendor's privacy policy and your admin settings to see whether training is on.

The Bottom Line

The AI meeting assistant has moved from stenographer to participant. That can save a real chunk of the hour after every call, and the math is easy to run for your own team. The consent rules have not caught up, so the tools that leave when an outsider arrives are the ones designed with that gap in mind. Try one on a single internal meeting, time the follow-up, and let the numbers decide.