The first wave of enterprise AI was essentially a better search box. An employee opens a chatbot, pastes in context, fires off a prompt, and gets an answer. Useful, but limited — productivity gains stay with the individual, not the organization.

Gabriel Hubert, CEO and cofounder of AI company Dust, argues we're now entering a more consequential phase: multiplayer AI, where agents stop being private tools and start functioning as shared participants in company-wide workflows.

What 'Multiplayer AI' Actually Means

The distinction Hubert draws is sharp. In single-player AI, "one person learns a better way to do something, but that improvement doesn't necessarily spread. The individual gets faster, while the company continues to work in roughly the same way."

In a multiplayer model, agents can "start participating in the same workflows as other people and other agents. They can hand work over, reuse what another team has learned and contribute to a shared system rather than starting from scratch every time."

In practice, this looks like a team-level workspace where users '@' mention specific agents — much like tagging a colleague in Slack — to perform discrete tasks, then pass results downstream to another agent or human. A 'blog writer' agent, for example, generates a draft and hands it to a 'LinkedIn' agent that produces social copy based on the same shared context. The workflow is versioned, improvable, and accessible to the whole team.

Hubert has also seen this applied in sales operations: rather than each rep spending 30 minutes per lead on research, CRM updates, and qualification — and doing it differently every time — a shared agent handles that entire process consistently. "The important part is that the workflow is shared, improves over time and becomes available to the rest of the team."

Governance Is the Hard Part

Scaling from individual tools to shared agent infrastructure creates serious governance challenges that most enterprises aren't ready for. According to data from Smarsh, while 55% of large EU enterprises are already using AI, only a quarter believe their governance models are fully equipped to handle it.

The gap opens the door to 'shadow AI' — employees using unauthorized tools outside the purview of IT, sometimes inadvertently surfacing sensitive data. If an agent has access to a company's Google Drive, for instance, any user interacting with that agent could stumble across confidential records.

At Dust, the solution is layered access control managed by administrators who decide which data sources connect to which agents and which teams can access them:

  • Some agent spaces are open company-wide; others are restricted to specific teams
  • Agents inherit the data permissions of the space they're built on — not the permissions of the user querying them
  • If a user lacks access to certain data, the platform blocks the agent from surfacing it to them

"The context and operating knowledge of a company should belong to the company. It shouldn't be trapped inside a model provider, an individual conversation or a series of disconnected applications." — Gabriel Hubert

The New Role That Makes It Work

Technology aside, Hubert believes the most important factor in multiplayer AI adoption is organizational: the emergence of the 'AI operator'.

This person isn't a data scientist or an IT admin — they're someone who looks at existing processes and asks whether those processes should still exist at all now that AI is available. Hubert describes their mindset as running an 'anti-to-do list': after every tedious recurring task, asking "how do I never have to do this again?"

Research from Microsoft backs up the organizational angle, showing that factors like manager support, talent practices, and culture drive more than twice the AI impact of individual employee effort alone.

What 2027 Looks Like

Hubert's near-term forecast is that by 2027, the debate won't be whether to use AI agents, but how to manage a workforce of them — tracking what they're doing, who's accountable, and whether their outputs remain reliable over time.

But he's careful not to reduce the equation to pure automation. As agents absorb more execution, he argues, human judgment becomes more valuable, not less.

"The agents themselves are not the entire asset. The asset is the loop between the agents, the context the company owns and the people who keep improving both."

For founders and operators building on top of AI infrastructure, that's the real strategic signal: the companies that win won't just be the ones that deploy the most agents, but the ones that build the best feedback loops between those agents, their proprietary data, and the humans steering both.