The problem isn't that AI agents can't do tasks — it's that they have no idea which tasks matter or why.
The Summary
- Zeroset raised $5.2 million from Gradient Ventures (Google's VC arm) to build models that understand company operations, not just company documents.
- Their product Nebula connects to enterprise tools like Microsoft 365, GitHub, and Teams to track how work actually flows through an organization over time.
- Two college dropouts, ages 18 and 21, are betting that context about workflow sequences matters more than raw document access for autonomous agents.
The Signal
Zeroset is solving the difference between knowing what happened and understanding why it happened. Current AI agents can search Slack, scrape Notion, and summarize emails. What they cannot do is tell you that when Sales closes a deal over $100K, Engineering always gets looped in before the contract is signed, or that Support tickets escalate to Product only after three failed resolution attempts.
Akshat Kannan, the 18-year-old CEO who left Stanford to build this, puts it directly: models off the shelf have no understanding of how an enterprise actually works. They lack the procedural memory that humans build by watching patterns repeat. A junior employee learns that you ping Sarah before touching the production database. An AI agent reads the security docs but has no idea Sarah exists.
"Models off the shelf have no understanding of the dynamics of how an enterprise actually works. They have no understanding of how the work gets done."
Nebula tracks information and activity changes across Microsoft 365, SharePoint, Outlook, Teams, GitHub, and workplace messaging platforms. The key word is "tracks." Not indexes. Not searches. Tracks how decisions cascade through systems. The goal is sequence awareness: when X happens, Y follows, and here's the evidence across 47 prior cases. That transforms an agent from a glorified search bar into something that can actually close loops autonomously.
This matters because the agent economy stalls without institutional knowledge. Companies have been digitizing documents for 30 years. What they have not digitized is the unwritten playbook of how work actually moves. That lives in email threads, Slack DMs, hallway conversations, and the heads of people who have been there long enough to know who to ask. Zeroset is building the missing layer: the workflow graph that connects documents to decisions to outcomes.
Three reasons this is harder than it sounds:
- Enterprise software is famously siloed. Integrating Microsoft 365, GitHub, and Slack simultaneously means navigating OAuth hell and rate limits.
- Temporal modeling is expensive. Tracking state changes over time across multiple systems requires more than embeddings in a vector store.
- Privacy and access control get messy fast. Not everyone should see everything, even if the agent needs it for context.
The $5.2 million round from Gradient Ventures (Google's AI-focused fund) signals that someone with deep pockets thinks this is solvable. Google has been building knowledge graphs for search since 2012. Applying that thinking to internal company operations is the logical next frontier. If agents are going to work for us, they need to understand us first.
The Implication
If Zeroset is right, the companies that win with AI agents will not be the ones with the most data. They will be the ones whose data is connected in time. The question is whether Nebula can scale beyond the closed research preview without getting crushed by incumbents like Microsoft, which already owns half the integration surface. Watch for customer traction in the next six months. If they land a reference customer in finance or healthcare, this becomes a real category.