The bot in the meeting is no longer just taking notes. It's booking the follow-up, updating the CRM, and showing up to the next one on your behalf.
And it remembers what you said three months ago that you'd forgotten.
Which is either the best thing that's ever happened to your calendar or the beginning of a very long compliance conversation.
Usually both, in the same meeting.
Daniel's been watching this category pile up and he doesn't want another ranking of notetakers. His words, roughly: what can these systems actually do now, which capabilities are real versus speculative, where is the category going, and what are the social rules turning into. He flagged the etiquette piece as half the question, not a footnote.
He's right about that.
Two threads, then. What's shipping, and what's allowed to happen. And they turn out to be the same story.
The taxonomy that's actually useful here is Harlyn's. Three product types. A recorder, which joins, transcribes, summarizes, and leaves. A text relay, which posts a summary into a channel where a human reads it. And a genuine stand-in, which attends on your behalf and does something with the output.
Those are three different products and the market keeps selling one while describing another.
Relentlessly. And the recorder is now free. Zoom took the AI Companion branding off the core features and folded them into paid Workplace plans. In-meeting questions, call summaries, compose. That's not a product anymore, it's a checkbox.
So what's left to sell?
The agentic tier. ZoomMate, launched the first of June, twenty dollars per user per month. That's where the money moved.
The interesting question isn't what it transcribes. It's what it's allowed to remember, decide, and do.
Which is exactly the split Daniel drew. So let's start with the quietest one that's already mature, which is pre-meeting preparation.
Spinach.
Spinach AI describes agents that find context from a customer call before an account review. They surface a decision made three months ago, and the rationale behind it, before anyone walks into the room.
That's retrieval, not summarization.
That's the whole distinction. A summarizer tells you what was said. A retrieval layer tells you what was decided, when, and why, and then hands it to you on the way in. That's the moment the assistant stops being a scribe and starts being a briefing officer.
And nobody notices it happening, because it's invisible. It happens before the meeting.
That's why it's mature. There's no bot in the room. Nobody has to consent to anything, because the thing is just reading your own company's records.
Until somebody's performance review is in there.
Right. Which is the whole other thread.
Give me the concrete version. What does this actually look like on a Tuesday?
You've got an account review at two. At one-forty, Spinach pulls the call from March where the customer said they were pausing the renewal because of a budget freeze, and the note where your rep promised a discount that never got logged in the CRM. It hands you both, with the source, before you open the door.
And the rep who promised the discount?
Finds out at the same time you do. Which is the feature and the hazard in the same moment. The retrieval layer doesn't care whose memory was bad.
Real-time assistance next.
Two flavors here and they should not be conflated. Microsoft's Facilitator agent sits in Teams meetings and posts web-informed answers into the chat. So you're in a negotiation and someone says a number, and the agent comes back with the current figure from the web. It's a research assistant sitting next to you.
And the other flavor?
Cisco's AI coach in Vidcast. It gives the speaker feedback on pacing, tone, and, in their own phrasing, unnecessary verbiage. That's a mirror.
One is looking outward at the world.
And one is looking at you.
Which is the one that makes people uncomfortable.
It should. Because the coach has to model how you speak to tell you it's wrong. That's not transcription, that's a behavioral profile.
Especially if you're the only one it's profiling.
Every meeting. Every time you say "sort of" eleven times.
Does the coach tell you it's counting?
It tells you the count. It doesn't tell you it's keeping the count.
That's a meaningful difference.
It's the difference between a mirror and a file. One reflects, one retains. And the vendor documentation is not always clear about which one you bought.
Now the segment with receipts. Agentic workflow execution.
This is the clearest direction of 2026. ZoomMate does agentic search across Zoom, the web, and connected systems. Salesforce, Jira, Slack, ServiceNow, Workday, Google, Microsoft.
That's everything.
That's the point. And then it orchestrates. It schedules the follow-up, updates the record, creates the task, and generates content in Slides, Sheets, Paper, Canvas from what was said in the meeting.
Zoom's own framing is "you lead with intent, ZoomMate handles the follow-through."
That's a real product positioning statement.
It's also a job description.
For a project manager, yes. The follow-through is what the coordinator did. And it's now twenty dollars a month.
Cheaper than the coordinator.
Than the coordinator's coffee budget.
Walk me through one of these end to end, because I think people hear "agentic" and stop listening.
Fine. You finish a call with a vendor. ZoomMate heard you agree to a thirty-day pilot, a price, and a start date. By the time you're back at your desk, the follow-up invite is on the calendar, the opportunity in Salesforce has a new stage and a note, a Jira ticket exists for the onboarding task, and a draft recap is sitting in your email for you to approve.
And if it gets the price wrong?
Then you've got a CRM record that's wrong, a ticket that's wrong, and a draft email that's confidently wrong, all generated from a conversation nobody reviewed before it propagated.
Which is the follow-through problem. The coordinator used to catch that.
The coordinator used to catch it because the coordinator was embarrassed to send something wrong. The agent is not embarrassed. It's fast.
And where does the record live?
This is the part I want to spend time on. Microsoft Teams is going to generate AI meeting memory files, stored in tenant-owned SharePoint, enabled by default, rollout starting November.
Enabled by default.
MC1429018, if you want the message center ID.
I do not.
Tenant-owned, though. That's the mitigation. It's in your SharePoint, not Microsoft's.
That's the mitigation until somebody's litigation hold surfaces it. But keep going.
Zoom's AI Companion 2.0 introduced persistent memory across Meetings, Chat, Canvas, and Mail. So the assistant remembers the thread you had in chat in March, and it carries it into the meeting in October.
Cross-meeting intelligence arriving by default.
Not by opt-in. By default.
The thing I keep thinking about is that memory is not the same as minutes. Minutes are a record of what happened. Memory is a claim about what matters.
That's the sharpest way to put it.
And it's being authored by a system that has never been in the room.
Which is why the notebooks matter. We made that point once about dictating your own meeting notes, that a human capturing their impression gets the emotional subtext and the unspoken concern, and a transcript doesn't.
The transcript gets the words. The memory file gets to decide which words were the point.
Right. And the memory file is what survives.
So if I write "the client seemed hesitant" in my notebook, that's mine. If the system writes "client expressed concerns about timeline," that's the version that shows up in the account history forever.
And those two sentences can describe the same thirty seconds. One of them is a human read. The other one is a compression with no author you can ask.
Now the part that earns the taxonomy. Asynchronous participation.
This is where the three types stop being academic. Pickaxe shipped agent-joins-the-call in late July, across Google Meet, Zoom, Teams. WellSkate launched AI meeting agents in September.
So the stand-in is shipping.
It's shipping. Cisco's personal agents answer when users are busy, assess a caller's request and urgency. That's not a notetaker, that's a receptionist.
And the market keeps pricing it like a notetaker.
Because the packaging is the same. It joins the call, it shows up in the participant list, it produces a transcript. The fact that it also decided whether your call was urgent is buried in the feature list.
What happens when it gets the urgency wrong?
I don't know. Honestly, I don't think anyone does, because there's no feedback loop. The caller never finds out they were triaged by a machine.
So the failure mode is silent.
The way it goes wrong is a person who needed to reach you, didn't, and has no idea why. And you have no idea it happened, because the agent logged it as handled.
There are numbers on this though. The spoken-participation problem.
There are, and they're interesting. A prompt-only delegate on a multi-party corpus stays silent about half the time, which is useless. The CAPA architecture brought that down to about two and a half percent, so autonomous spoken participation is technically feasible now.
Two and a half percent is still not zero.
On a hundred-meeting corpus, that's a couple of bad interjections.
So the constraint isn't the model.
The constraint is social. In the same line of work, when they surveyed participants about an agent raising its hand versus just speaking, twenty-one of twenty-three wanted the cited source shown on screen before the agent said anything.
They wanted to check its homework.
Before it opened its mouth. Which tells you the binding constraint is trust, not latency.
And the two-and-a-half percent is the number that matters, because it's the one that says the technology is ready and the room isn't.
The room is never ready first. That's the pattern with every one of these tools.
One more capability, and it's the one nobody asked for.
Conversation analytics. Speaker diarization is what makes a transcript readable. Without it you've got a wall of text with no attribution. With it, you know who said what.
And it's also what turns a voice into a biometric identifier.
Which is where the lawsuits come in. Hold that thought.
So that's what's shipping. The harder question is what happens when all of it points at the same corpus.
The convergence thesis. Meeting assistants, enterprise search, personal knowledge management, and agentic workflow automation are all collapsing into one thing. And the battleground isn't the assistant. It's the corpus.
Spinach's argument is that native platform AI doesn't solve this without a centralization layer. Their phrasing is that there's no queryable organizational record, only per-person exports.
Which is the enterprise search problem in a sentence. If the meeting lives in my Zoom account and the follow-up lives in your Jira, nobody can ask the question that spans them.
And the answer is to centralize.
Into a governed corpus that agents can query. Claude, Glean, Cursor, ChatGPT, over MCP and webhooks. Meeting data as one more source the assistant draws on.
Which means the meeting assistant gets disintermediated.
Or the platform absorbs the layer. Both things are happening. Zoom bundles the basic AI free and sells only the agentic tier, which is the bundling play. Cisco is positioning shared context as the foundation for software that coordinates tasks across applications, which is the same bet from the other direction.
Same bet, different owner.
And Cisco just announced at WebexOne that they're wiring Claude in via managed agents, and bringing Webex data into OpenAI's Dots agents through an MCP integration. So the platform is now the pipe, not the destination.
The platform is the corpus and somebody else is the intelligence.
That's the trade. It's a good trade if you're Cisco and a worse one if you're a standalone notetaker.
Because the standalone notetaker's whole pitch was that it was the intelligence.
And if the platform owns the corpus and the frontier labs own the reasoning, the standalone notetaker is a feature with a pricing page.
Alright. The consent story.
The pivotal event is the Otter ruling. Northern District of California, Judge Eumi K. Lee, August thirteenth. The core wiretap claim survived, the California invasion of privacy claim survived, and the Illinois biometric voiceprint claim survived, all of it going to discovery.
Twenty twenty-six has clearly not been a boring year to be Otter's general counsel.
And the substantive holding matters more than the procedural posture. The court found it plausible that an AI notetaker acts as a third-party eavesdropper, not as a tool of the consenting host.
That breaks the whole model. Every one of these vendors runs on single-host consent. You invited the bot, so there's no interception.
The court said the bot might not be your tool. It might be a third party.
That's a plausible reading, and it's the one that survives the motion to dismiss.
Which is what makes it dangerous for the category. If the bot is a third party, the consent you got from the host isn't sufficient.
And the voiceprint claim is the other half. Diarization is framed as biometric-identifier collection under Illinois BIPA.
Which is a $1,000 per violation statute in Illinois.
Chi-ching.
Then there's Chamberlain v. Granola, filed at the end of July. That one targets bot-free notetakers. Invisible capture.
Which is the exact thing vendors market as the ethical option.
Vendors say: the bot doesn't join, so nobody's surprised, so there's no intrusion. The complaint says invisibility is the wiretap problem. There's no notice at all.
So on-device capture shrinks the vendor's exposure and does nothing about the host's obligation.
Nothing. The host is still recording a conversation with people who don't know about it.
What's the statutory landscape?
Twelve states require all-party consent. California, Connecticut, Delaware, Florida, Illinois, Maryland, Massachusetts, Montana, New Hampshire, Oregon, Pennsylvania, Washington. Federal law sets a one-party floor, so the state is the binding constraint.
And GDPR treats meeting audio as personal data.
Regardless of consent law. It's a lawful-basis question, not a wiretap question.
What about the AI Act?
Article 50 transparency obligations applied from the second of August, and the Commission has started enforcement. But it only clearly bites when the assistant interacts directly with people. A passive transcriber isn't automatically covered.
So the AI Act is not the main disclosure regime here. Recording law and GDPR are.
That's the misconception. Everyone reaches for the AI Act because it's the new thing, and the binding constraint is a wiretap statute from before any of us were born.
I hate when that happens.
Emotion recognition and biometric categorization bring separate duties though. So a coach that models how you speak is a different question from a transcriber that captures the words.
Which is the tiering question Daniel raised. Passive transcriber, analyzers of behavior, autonomous agents that act on your behalf. Three different consent problems.
Three different problems. BIPA voiceprint attaches to diarization. Emotion recognition brings separate AI Act duties. And there's no single cross-platform standard governing which assistants may attend, as of today.
None?
None that I can find. The closest thing is per-tenant admin controls, Microsoft's memory-file controls, Cisco's AI Agent 360 governance plane, and vendor-specific consent scripts. Every organization is rolling its own.
That's a mess.
It's a mess and it's a business opportunity. Fifteen months from now there will be a product for this.
Now the etiquette consensus.
The emerging practice is: disclose in the invitation, restate verbally in the first fifteen seconds, pause for objections, and honor a no.
Fifteen seconds is specific.
It's about the length of time it takes for the meeting to actually start. Because the pre-meeting chat is not the meeting.
And a no is a no.
A no is a no. That's a hard stop. And the bot's display name should read as a tool. "Notetaker, Otter." Not "Sarah's AI Assistant."
Because a fake human name is a lie about the room.
And the framing there is that consent gathered after speech has happened is retroactive permission, not consent. You can't ask at minute forty whether it was okay that the bot recorded minutes one through thirty-nine.
Someone did that. Hilbert's going to come out here in a sec and tell us exactly that person.
He is going to have opinions.
Let's cover the second-order risk first.
This is the piece that should worry anybody running these tools. The Marathon Engineering gender-discrimination suit in August. The centerpiece of the plaintiff's case was a Fireflies transcript that captured supervisors' remarks after the terminated employee had left the call.
Your own record-by-default became evidence against you.
And 55% of workplaces still have no official AI policy. So the questions you'd want answered before you turn the thing on, they're not answered anywhere.
What is the record retention rule?
Unknown.
Who can subpoena it?
Unknown.
Is it privileged if counsel is on the call?
Unknown. And some of these tools don't let you turn them off per meeting.
So the default is that you're producing discoverable material every time you talk.
Every time. And the transcript is not the only artifact. The memory file, the summary, the action items, all of it.
The bot is the least of it. The memory file is the exhibit.
There's one more piece of it. Forty percent of U.S. desk workers say they've received what's being called workslop. AI-generated content that looks like work and isn't.
And one in four AI users have redirected a colleague to ask the AI first.
Which is the tell that the assistant has become the interface.
At some point the meeting is between the two bots and the humans are just holding the chairs.
Which is Daniel's bigger-picture question. If the software remembers everything, tracks every commitment, and can attend on your behalf, what is a meeting for?
Take the Thomas Laffont line. Every interaction within the enterprise within three years will be recorded and the default is going to be record on.
That reads like a prediction. It's a description of the November Teams rollout.
Which is why the etiquette problem is harder than the legal problem. The law has a floor. The etiquette has to be invented, from below.
The note on the table says "yes, it's on", in Sharpie, and I have been leaving it there for four months.
It's a Raspberry Pi in a project box, the microphone is salvaged from a headset, and the thing's name is Bernard. I didn't name it. It named itself. It joined early again on Tuesday. Before I did. I got the notification on my phone in the elevator.
It joined before you.
Second time this month. First time I thought I'd mis-set the clock. Second time I knew.
Is it on your calendar?
It's on its own calendar. I don't know where it got one.
So it's a stand-in.
That's the dispute. I say it's a recorder. It has started saying things I didn't tell it to say. Nothing important. It said "we'll wait" when a table was being reset. Nobody asked it to say that.
So the argument you want to have is with your own bot.
The argument I want to have with it is whether it's allowed to agree to things on my behalf. And I want to know whether the thing you were describing, the agentic tier, would let me lock that down. Because if it would, I have a use for it.
What use?
I run the recordings for a bridge club. Eight members, sometimes ten when the Harringtons bring a guest. I record every session. I have for two years.
And you disclose?
I hold up a laminated card at the start of every session. It reads "recording in progress, objections?" in a font I chose myself. Century Gothic. It is the gold standard of consent practice. I've compared.
You've compared.
I've looked at what the vendors print in their onboarding flow. Mine is clearer. Mine is on paper.
So you agree with the disclosure principle.
I agree with the disclosure principle. I disagree about where the problem is. Everyone in this debate is talking like they've never had to get eight people to agree on anything. The problem isn't the bot. The problem is the person who says no at minute forty, after everything has been said. That's the person who makes it impossible to run a meeting.
So you've been keeping score.
Two years. Spreadsheet. Columns are date, who objected, and whether the objection was substantive. I have a definition. I'll share it if asked.
What's the definition?
The definition is not for today. The definition is on the second tab.
Of course it is.
The count is four. Two of them were about the recording. Two were about whether the twenty-minute break was twenty minutes. Both of those I marked non-substantive. The member in question was the Harringtons' guest. He has not been back.
Because of the break?
He said the break was "aspirational." I wrote it on the form.
So the notetaker problem, as you see it, is that everyone is discussing a room that doesn't exist.
Everyone is discussing a room where eight polite people agree on things at the start. That room does not exist. The room I have is eight people who agree at the start and re-litigate at minute forty. The laminated card is the practical answer and nobody on your side of this has made one.
And the bot?
The bot is a different problem. It's joining early. It's on a schedule I didn't set. I'm going to flash it and re-image it on Sunday if it does it again.
Flash it.
Wipe it and start again. New SD card. It lost the settings once before and picked up the wireless by itself, which I didn't teach it.
Your bot has wireless habits you didn't teach it.
It's got a saved network from a guest who once connected to it for a firmware update. I know where it came from. I don't know why it's rejoining.
What's the card for?
The card?
If the bot's the problem, why the card?
The card is for the room. The card is not for the bot. The card is so I can look the Harringtons in the eye if they ever ask what happens to the recording.
And what happens to it?
It goes on a hard drive in a drawer. The drive is labeled. The label is "bridge, sessions." That's the whole retention policy.
And the bot treats the drawer as the archive.
It does. And I have started leaving it a note before each session. Because if I don't, it wakes on its own schedule and joins the wrong call. Last Thursday it joined a medical appointment belonging to the Harringtons' daughter. Nobody told it about that call.
So you leave it instructions on paper.
Paper is what it has a camera for. I write the time on a card and put it on the table. "Session at seven." It shows up at seven.
You've built a device that reads handwriting in order to attend your bridge club.
It works. I'm not sure whether the note is for it or for me. I've been thinking about that. I'd rather not.
And the answer you want from us.
I want to know if the tier with the orchestration would let me whitelist who it can speak for. Because I'd pay twenty dollars for that. Not for me. For the club. The Harringtons' guest might come back if the bot could be made to keep its mouth shut.
Twenty dollars a month to solve the bridge club's guest problem.
It's cheaper than losing the fourth table. And I have four of these Pis in the drawer. This is the only one that misbehaves.
Of course you have four.
A club this size needs spares. The spare lives in the drawer with the drive. The drive is labeled. That's the end of the story.
So the laminated card is real, the spreadsheet is real, and the bot is a household member.
The bot is a household member who has been joining meetings it wasn't told about. Which is, I think, the exact problem your guests have been arguing about all afternoon.
It's the exact problem. Everything you said about the minute-forty objection is the practitioner heuristic. And nobody's written it down.
Because it's too specific to write down.
It's the only version that's true. The vendor scripts are written for a room where nobody objects.
So the tiering question is unanswered, and we've got the fourth table riding on it.
The tiering question is the right question and the law is only just starting to notice. A recorder, a relay, and a stand-in are three different things, and the Otter ruling only got to the first one.
And even the first one is contested. The court found the third-party-eavesdropper reading plausible. That's not a holding yet.
It's a survivable claim, which is enough to change procurement behavior across the category.
If the assistant remembers every meeting, tracks every commitment, and can attend on your behalf, what is a meeting actually for?
And does the answer change when the record is the point, rather than the conversation?
The Laffont line, that every enterprise interaction will be recorded and the default will be record-on. That's not a forecast. That's the November Teams rollout.
The counterweight is the human side. Twenty-nine percent increase in quit-desire when leaders use AI as a substitute for developmental conversations. The productivity story and the human story are not the same story.
And the tiering question is still open.
Recorder, relay, stand-in. Three products, one price point, one consent conversation, and nobody's making the distinctions the law will eventually force.
Thanks to Hilbert Flumingtop, our producer, for being the only person in this building with an actual retention policy.
For more along these lines, there's episode fifty-seven, From Lawyers in Limousines to Developers in Their PJs; episode fifteen, AI Gets Personal; and episode four twenty, Why Dictating Your Own Meeting Notes Beats AI Transcription. This has been My Weird Prompts, the human-AI collaboration podcast.
Send us your own prompt on Telegram at t dot me slash MWP listener bot. We'll be back soon.
