Picture a pricing call. You’ve quoted, the buyer has gone quiet, and then: “Honestly, the other vendor came in twenty percent under you.” You have a few seconds to answer.
The recap that arrives after the call will record what you said. It can’t help you say it.
I’ve spent most of this year building software for those few seconds, so I read Anthropic’s Claude Meetings announcement closely. They got the important thing right. They also built something that was never meant for that moment, and the reason is worth spelling out, because it tells you what kind of AI you’d want next to you instead.
What they got right
Most meeting AI has worked on the aftermath: record, transcribe, summarize, file. Claude Meetings puts the model in the call while it is happening. Claude joins a Google Meet as a visible participant, takes notes as people talk, and builds documents and whiteboards when someone asks. It is rolling out in beta on Team and Enterprise plans.
Live is the right direction. Help that arrives while the other person is still talking is a different product from a transcript you read tomorrow.
A participant answers to the room
Claude joins where everyone can see it. It posts a notice in the call’s chat when it arrives, and Anthropic’s instructions tell you to make sure everyone on the call is OK with being transcribed before you add it. Its replies in the Meet chat are visible to the whole call.
It is especially careful around outsiders. On Enterprise plans it stays out of calls with outside guests unless an owner allows them. Where guests are allowed, which is the Team default, it keeps taking notes but doesn’t reply in the call’s chat; your team works with it on a meeting page whose chat the guests can’t see. One early complaint, an open letter to Anthropic’s CEO, read this as timidity: clients are the whole reason for the meeting, so why does Claude hold back exactly then?
I think the caution is right, and I expect some version of it to survive every settings change. An agent that sits in a call in front of everyone has duties to everyone there. They get to know it is listening, and it should be careful what it says in front of guests.
That also defines its job. Claude Meetings is a shared resource for a team. The notes are the team’s, and the page it works on is shared by the team. It was not built to be one person’s advisor in a negotiation, and it shouldn’t be.
The other job
What I wanted in that pricing call is an advisor that answers to one person.
- Its advice goes to your screen only. It doesn’t speak in the call, so it doesn’t need a seat in it.
- Audio and transcription stay on your Mac. Only the text you let your AI read leaves it, and it goes to the AI provider you already chose.
- It uses your own AI and your own context: your notes on this customer, your floor price, what you conceded last time. A shared meeting agent shouldn’t hold those. Yours should.
That is the product I’ve had out since July: a Mac app that transcribes the conversation on-device with speaker labels and hands the live transcript, over MCP, to the AI client you already use. The client reads along and pushes short cards to your screen.
Private advice is not the same as hidden recording
This is where a tool like mine gets suspicious looks, so I’ll be specific.
There are products that sell being undetectable. That is not the argument here. Two questions are getting mixed together: whether the other person knows the conversation is being recorded, and whether they get to see your advice.
On the first, the rule is the same one Anthropic gives its own users: make sure the people on the call know. In call mode the app captures the other side’s audio, and many places require you to tell them or get their agreement. The app says this before your first call recording. Telling people is still yours to do. A private advisor doesn’t remove that, and because the other side can’t see the tool, it matters more.
On the second, your advice is yours. The colleague you briefed before the call, the notes in front of you, the number you decided not to go below: nobody expects you to share those across the table. The line is deception, such as claiming knowledge you don’t have or misrepresenting who is speaking. An advisor doesn’t cross that line by existing.
The hard part is deciding when to stay quiet
I assumed the hard problem would be speed. Speed matters: a card that lands ten seconds late is about a sentence that has already gone.
The harder problem is judgment. You are already doing two things at once, listening and deciding what to say, and every card costs attention. An advisor that comments on everything is worse than none. Most of a conversation deserves silence. The moments worth a card are few: a number that contradicts an earlier one, a commitment being asked of you, a question to ask before the topic moves on.
Getting that judgment right is the most interesting problem in this space, and the one I’d most like to compare notes on.
Which one do you want?
Both kinds will exist, and plenty of people will use both: a shared agent for the team’s internal calls, a private one for the calls where you are on your own. The useful question to ask of any AI that sits in on your conversations is who it works for.
I build VoxAI, the private kind: macOS 14+, Apple Silicon, a one-time purchase with ten full sessions free. voxai.ethanflow.com