By Hamza Faisal
Two college dropouts have raised $5.2 million to fix one of the most stubborn problems in enterprise AI: getting autonomous agents to actually understand how a company works.
Zeroset, a San Francisco startup co-founded by Akshat Kannan and William Zhang, emerged from stealth this month with a pre-seed round co-led by Gradient Ventures and 2048 Ventures, with participation from Leblon Capital.
Kannan, 18, left Stanford University earlier this year to run the company as chief executive; Zhang left the University of Texas in 2025.
Their bet is that increasingly capable AI models still fail at enterprise automation because they lack an evolving picture of the organizations they operate in — the unwritten processes, dependencies and decisions that never make it into a company wiki or a training manual.
Nebula, the "state layer" for work
The startup's first product, Nebula, connects to workplace systems such as Microsoft 365, Outlook, Teams, SharePoint and GitHub, capturing the operational traces that people and AI agents generate as work happens.
From those traces it builds models of how the organization functions and changes over time — what the founders call enterprise "worlds."
That is a deliberate contrast with the retrieval-heavy approach most enterprise AI tools take today.
Nebula is not designed to fetch an answer when an agent asks a question; it models the flow of work itself, creating a persistent representation that helps agents behave consistently over time rather than merely answering queries.
"Models off the shelf have no understanding of the dynamics of how an enterprise actually works," Kannan said.
He argues the gap is why so many early enterprise AI deployments stall after their first wins — an agent that answers customer emails fine can still flounder when a workflow requires coordinating across Slack, GitHub and a CRM, losing track of decisions made along the way.
A robotics lesson, applied to the office
Zeroset draws an explicit parallel with robotics, where increasingly capable robots needed not only better policies and more computing power, but representations of their surroundings and an understanding of how those surroundings changed in response to their actions.
The founders want to apply the same idea to businesses: learning models of the enterprise world from observed traces of real work.
Under the hood, the company describes a custom hierarchical vector graph that maps how work moves through an organization, turning observed patterns into what it calls repeatable playbooks.
Nebula is currently in a closed research preview with access expanding gradually; the company has sketched hybrid pricing — a license fee plus usage charges tied to data volume and agent activity — but has not announced public pricing.
A small round in a crowded battlefield
The round is modest by 2026 AI standards, and the startup declined to disclose its valuation.
But the ambitions are not small: Kannan said a single peak model-training run cost about $100,000 over two weeks, and the five-person team plans to use the money to hire researchers and engineers, train its models and support early enterprise deployments.
Zeroset enters a fiercely contested market, with a generation of agent-infrastructure companies racing to become the memory, orchestration or control layer for AI workers inside companies.
