Ask an AI agent to draft an email recapping your latest internal sync, and it won’t break a sweat. Ask it to find you the cheapest flight to Denver, and you’ve got a shortlist in seconds. But ask it to carry a product request through to release, figuring out whether last week’s Slack approval still applies to this week’s changes, and it will likely stall.

The models behind today's agents have absorbed most of what the public internet knows about tasks like drafting an email or booking a flight. The prevailing assumption is that enterprise automation is now a waiting game: the models keep improving, and agents will eventually work their way into every company.

That assumption is flawed: no model, no matter how capable, knows the ins and outs of how a company operates. It does not know that a claims analyst at one insurer checks an internal financial dashboard, reads the related email thread, and then updates a record in the CRM, in that exact order, with exceptions that live in nobody's documentation. That knowledge is spread across dozens of private tools and the daily habits of the people who use them. Until it is captured somewhere an agent can use, enterprise agents will keep stalling a few steps in.

Akshat Kannan and William Zhang are building Zeroset, which recreates workflows and states of how enterprises run so agents can operate inside it. Today, Zeroset announced a $5.2 million pre-seed round co-led by Gradient and 2048 Ventures, with participation from Leblon Capital.

Why do agents stall inside the enterprise?

Consumer agents achieved remarkable results so quickly because they were trained on endless amounts of information: the public web. Every search, checkout, and booking flow has been documented and rehearsed millions of times, so models had the luxury of learning those tasks from an abundance of examples. Enterprise work is the polar opposite. The ticketing system, the internal dashboard, the approval flow, and the CRM are all private by design, and the sequence that connects them exists only in the minds and click histories of employees.

Big tech companies have started recording how their own engineers move through internal tools, with the goal of automating the work around the code: testing, requirements review, deployment. And they can afford to build this in-house, virtually nobody else can. The enterprise desperately needs a vendor whose job is to organize those traces into something an agent can act on.

What Zeroset builds

Zeroset starts with the traces: the historical clicks, steps, and handoffs employees leave as they move through their tools, plus the documents and records that give those actions context. Its system connects them into a live representation of business state: what happened, what changed, what is still unresolved, and how the work of one person relates to the work of another. From those relationships it finds the patterns and hands agents a working model of how each workflow runs at this company, for this role.

Its first product, Nebula, is now in closed research preview. Long-running agents use Nebula to query the current state of a workflow rather than reconstructing that context from raw data and history every run. On LongMemEval, Nebula delivered 20% higher retrieval accuracy than Mem0, Supermemory, and naive RAG, while returning results in under 50 milliseconds at the median and using fewer tokens. On the BPI Challenge process-mining datasets, a Nebula-based system outperformed all recorded implementations on next-activity prediction by more than 10%, testing a harder capability: anticipating what happens next in a workflow.

What changes when agents can see the whole workflow

There’s both an art and a science to understanding how those records fit together. A PM can check whether the latest production build meets the agreed requirements. Deciding whether a loose end should hold up the release, or whether a “looks good” in Slack counts as approval, takes judgment built through working with the team. An agent carrying the request through to completion needs that context alongside the formal requirements, plus a way to keep track as things change. For long-horizon agents, a state layer preserves the decisions already made and the circumstances behind them, helping the agent judge what should happen next.

Post-training an agent on a company's workflows requires those workflows to be captured first. Deploying an agent into them requires the same. Zeroset is the prerequisite for both, upstream of everything else in the enterprise agent stack.

Who owns the workflow data

In the future, we expect the companies that have gotten agents to work will share one common trait: a central record of how their work actually gets done, across every tool, by every employee and every agent. That record is the most valuable dataset an enterprise has, and it should stay inside the enterprise. A company that holds its own workflow traces can use the best frontier model for execution without handing over the knowledge of how it operates, and it can train open-source models on that knowledge if it chooses. Independence in how you deploy agents starts with owning the data they learn from.

Frontier models can already execute tasks. What they cannot do is understand a business: its state, its workflows, its unwritten rules, and how each of those changes over time. Akshat and Will are solving that at the infrastructure level, and they are hiring researchers and systems engineers to do it. If your agents keep stalling at the edge of your own systems, you should be talking to Zeroset.