The Secret AI In Your Breakout Room

You're ten minutes into your company's quarterly virtual town hall meeting with hundreds of other employees, sorted into a breakout room of eight coworkers arguing over whether the office should go fully remote. Sarah from finance makes her case. Dave disagrees, loudly, from his kitchen.

Then someone new joins. You don't recognize her. She says she's been listening in on a few other rooms, and drops a line that stops the debate cold. Three other groups raised a headcount-cost argument no one here had thought of.

It's a good point. You find yourself nodding.

The new participant is synthetic, stitched together from summaries an AI pulled out of seven other breakout groups, sent into yours wearing a friendly, invented face. As far as your Zoom window is concerned, she's just the new person in Room 4.

A recently authored patent describes exactly this. One giant conversation across a thousand people who aren't talking to each other, carried between rooms by AI-generated colleagues. The question worth sitting with is whether you've just been outvoted by an algorithm wearing a nametag.

Unanimous A.I. based in the U.S. filed "Scalable Methods and Systems for AI-Facilitated Video-Conferencing Among Large Conversational Human Groups." Behind it is Louis Rosenberg, a Stanford-trained engineer who built one of the first working augmented reality rigs at the Air Force Research Laboratory in the 1990s and has spent the past decade chasing what he calls collective superintelligence.

Video calls fall apart past about seven people. Everyone has sat through a fifty-person Zoom call where three people talk and forty-seven watch. Turn-taking collapses long before the video ever buffers, people talk over one another, and the group gets too big for people to track who's replying to whom.

Unanimous's fix is to stop pretending the group is having one conversation and admit it's many. The patent splits a large population, examples run from dozens up to 2,000 people, into small breakout rooms of five to ten, each running its own live video, text, or voice chat.

The new part is what happens between them. Each room gets a "conversational observer agent," an AI (the patent points to large language models like ChatGPT) that watches the dialogue every few minutes and scores how strongly people seem to believe the key arguments raised. That summary goes to a second AI, a "conversational surrogate agent," assigned to a different room. It joins that room's conversation in the first person, complete with its own username and, in the video version, an animated face and voice built on tools like D-ID's talking-head generator.

Do that across every room and information moves through the whole population without anyone leaving their small group. The patent's own worked example involves 800 football fans split into 80 rooms of 10, arguing who wins the Super Bowl, each room fed what the other 79 concluded until the crowd converges towards one answer.

What changes the stakes is that the messenger relaying other rooms' views is never labeled as a messenger. They look like a legitimate participant.

Large groups can't deliberate effectively. The usual workaround is a trade. Shrink the group and you keep a real conversation but lose most of the population. Keep the population and you're running a survey and losing productive discussion.

The real bottleneck is timing. Nobody has found a way to hold one honest, large conversation and get a cohesive answer back quickly and productively.

Extra! Extra! Read All About It!

This category sits between market research, workplace collaboration, and civic tech, and nobody owns it outright yet.

The closest direct match is Remesh, built to run live, AI-moderated conversations with up to a thousand people for clients like Deloitte and Barclays. Remesh clusters and scores responses as they arrive rather than routing them through synthetic go-betweens. It optimizes for one moderator reading a crowd. Comparatively, Unanimous's patent optimizes for the crowd talking to itself.

Adjacent to that sits Google DeepMind's Habermas Machine, a research system published in Science in October 2024 that mediates political deliberation by drafting consensus statements and revising them against feedback. Tested with more than 5,700 UK residents, it was rated fairer and clearer than human mediators. DeepMind's version writes one group statement. Unanimous's puts a synthetic avatar directly into the conversation as a peer, a more invasive design with fewer guardrails so far.

Further out is Pol.is, the open-source consensus-mapping tool Taiwan's vTaiwan process used to help write ride-sharing regulation, now being extended with LLMs partly through a 2023 OpenAI grant program. Pol.is clusters opinions instead of manufacturing dialogue, more transparent but less conversational.

The nearest paying market is AI meeting assistants, including the Otter.ai and Fireflies.ai layer that already summarizes ordinary video calls. Grand View Research puts that category at US$3.47 billion in 2025, growing toward US$21.5 billion by 2033. Incumbents have the distribution, but none of them yet do cross-room synthesis at scale.

Widen out and there's market research, an industry PR materials pegged at US$71 billion as far back as 2018, and civic or workplace deliberation, where governments and large employers now experiment with AI-run town halls instead of quarterly surveys.

None of that money flows to Unanimous directly. It shows real capital behind the idea that aggregating what large groups currently believe, in real time, is worth billions when the belief in question is worth it.

The company's leverage is timing rather than distribution. They have a decade's head start on pooling group intelligence, with 40 patents dating as far back as 2015, and a working product since 2023, while the giants are still bolting summarization onto video calls one update at a time.

Remesh, the closest direct competitor, has raised a total of US$42.7 million since 2018 and now serves more than 600 clients including Deloitte and Barclays.

Google DeepMind's Habermas Machine is a research project rather than a company, but its October 2024 publication in Science, backed by a 5,700-person study, is the biggest validation this category has had. OpenAI put smaller, earlier money behind a related idea, funding ten teams building on the open-source Pol.is platform with US$100,000 grants apiece in 2023.

The video layer this patent leans on is its own gold rush. D-ID, the digital-human company the patent names directly as an example, has raised a total of US$48 million, most recently a US$25 million Series B led by Macquarie in 2022. Synthesia closed a US$180 million round in January 2025 at a US$2.1 billion valuation, up from US$1 billion less than two years earlier, and HeyGen raised a US$60 million Series A that valued it at US$500 million and now reportedly generates close to US$95 million a year. None of these companies do cross-room synthesis. They build the faces and voices that make a synthetic colleague believable in the first place.

Zoom is still shopping, too. In July 2026 it agreed to buy Common Room, a Seattle startup that had raised US$52 million, to bolt AI buyer-intelligence onto its platform. Common Room has nothing to do with cross-room deliberation, but the deal is a reminder of how fast Zoom moves once it decides it wants a new AI capability, the exact dynamic Unanimous is racing against.

The real risk is that the synthetic messenger relaying information between rooms is never identified as one. Every surrogate agent in this patent is designed to pass as human, with an ordinary username and, in the video version, a face, a voice, and simulated inflection built on tools like D-ID's talking-avatar technology. The patent is explicit that in some configurations the agent shouldn't identify itself as a summary of another room. It should just speak in the first person, as if it had been there.

That's a deliberate design choice, and it collides with regulation arriving almost the same moment this patent issued. The EU AI Act's Article 50, effective August 2026, requires systems that interact directly with people to disclose that they're AI, unless it's obvious from context. California's SB 243, in force since January 2026, requires the same whenever a reasonable person could otherwise believe they're talking to a human. A synthetic colleague with an invented face in a corporate town hall sits close to what both laws were written for.

The patent also lets a human moderator overweight, underweight, or bar certain positions from being shared between rooms, a real content-moderation feature and also a lever for shaping what a thousand people believe the group concluded, with none of them able to see the lever being pulled.

The whole architecture assumes that letting rooms hear each other makes the group smarter, but convergence and correctness aren't the same thing. Google DeepMind found something similar with its own Habermas Machine, where AI-mediated groups converged faster, but support grew more for majority positions than minority ones, even when the AI dutifully worked minority objections into its summaries. A room that hears "most other rooms think X" is often responding to social proof more than to a stronger argument.

That distinction matters more in civics than in commerce. A crowd converging on one Super Bowl pick is the entire point of a forecasting exercise. A citizen assembly, a union vote, or a company town hall converging on one answer because an early opinion happened to spread first and get repeated by a synthetic peer in every other room is closer to the cumulative-advantage effect researchers found in the classic Music Lab experiment, where arbitrary early popularity snowballed regardless of underlying merit.