AI has learned to chat, code, and summarize with startling fluency, but it still stumbles when groups of people need to work together. Humans&, a three-month-old startup formed by veterans of Anthropic, Meta, OpenAI, xAI, and Google DeepMind, is betting that coordination — not conversation — is the next breakthrough. The company has raised a $480 million seed round to build a foundation model explicitly tuned for social intelligence and long-horizon teamwork, aiming to become the connective tissue across organizations.
Why Coordination Is the Real Bottleneck for Teams
Most large language models optimize for single-turn correctness and user satisfaction. That’s useful for one-off answers; it’s mismatched to the reality of ongoing projects, shifting constraints, and multiple stakeholders. Coordination requires skills today’s models rarely exhibit: eliciting preferences, reconciling trade-offs, tracking commitments over months, and facilitating decisions that stick.
There’s mounting evidence the market is hungry for this layer. McKinsey’s latest State of AI report found roughly 65% of organizations are using generative AI in at least one function, yet many cite workflow fragmentation and change management as the drag on impact. Reid Hoffman has argued that the next wave of productivity comes from AI embedded at the workflow level — how teams share context, run meetings, and make calls — not from isolated pilots that churn out documents no one reads.
That view is translating into funding beyond Humans&. Note-taking startup Granola recently raised $43 million at a $250 million valuation to ship more collaborative features, and major platforms are pushing “cowork” paradigms rather than standalone chat. Still, none of the incumbents are rebuilding their core models around social reasoning — a gap Humans& wants to exploit.
Inside the Model Blueprint for Coordination AI
Humans& describes its system as a “central nervous system” for collective work. That implies three ingredients beyond typical chatbots: memory, planning, and multi-party negotiation. Instead of optimizing for immediate thumbs-up ratings, the company says it will train for outcomes over time — did the group converge faster, did the plan hold, and were trade-offs made explicit and fair?
Technically, that points to long-horizon reinforcement learning and multi-agent training. The research community is already circling these ideas: DeepMind’s work on multi-agent RL, OpenAI’s orchestration frameworks for tool-using agents, and a growing body of academic benchmarks for multi-step collaboration. Expect Humans& to pair this with persistent memory structures — think profiles of each participant’s goals, constraints, and past decisions — so the model can ask fewer, better questions and surface disagreements before they derail a project.
Consider a mundane but costly example: choosing a new brand identity. Rather than collecting random opinions, a coordination model could elicit criteria from design, legal, and sales; weight them; propose candidate options; record rationales; and schedule decision gates. The success metric isn’t a clever quip — it’s whether the team aligned in fewer cycles with higher adoption.
Product Will Co‑Evolve With the Model and UX
Unlike assistants that plug into Slack or Google Docs, Humans& plans to own the collaboration layer outright. The team says the model and interface will be designed together, allowing behaviors like role assignment, decision logging, and escalation policies to be native, not stitched-on bots. That could look like a multiplayer canvas where the AI manages agendas, synthesizes positions, and updates a shared memory the moment a decision is made.
The company is targeting both enterprises and consumers, framing the product as useful for a 10,000-person org or a family coordinating budgets and calendars. If successful, it becomes the system of record for “who decided what, why, and what happens next” — a layer most companies currently cobble together across email threads, docs, and chat.
A Crowded Race With Heavyweights in Workplace AI
Humans& is stepping into a contest with giants. Anthropic is pushing Claude Cowork to shape team workflows, Google’s Gemini is embedded across Workspace, and OpenAI is pitching multi-agent orchestration to developers. Those players control distribution inside tools teams already use, and they are rapidly improving meeting notes, document drafting, and task handoff.
The differentiator for Humans& is the willingness to re-architect around social intelligence rather than bolt coordination onto a general-purpose model. That either gives the startup a head start on a new category — “coordination-native” models — or makes it an attractive acquisition target for platforms that want to deepen collaboration without tearing up their stacks.
Risks and What to Watch as Coordination Scales
Training a new foundation model is capital- and compute-intensive, and access to leading-edge chips remains constrained. Beyond infrastructure, the hardest problem may be evaluation: proving that teams using a coordination model consistently ship faster with fewer errors. Expect Humans& to pilot with design partnerships and publish metrics that look more like operational KPIs — cycle time, rework rate, decision latency — than chatbot leaderboards.
Regulatory and ethical questions loom, too. A system that maps team preferences and power dynamics must safeguard privacy, honor consent, and avoid entrenching bias in “who gets heard.” Organizations like Stanford HAI and the Cooperative AI Foundation have urged rigorous testing for cooperation and fairness; any vendor claiming to be the nervous system of work will face scrutiny here.
If Humans& can demonstrate durable gains on real projects — not just slick demos — the company could define a new era of AI that helps groups think together. The breakthrough many teams want isn’t a smarter chatbot; it’s a reliable partner that remembers, mediates, and moves everyone forward.
