Skan AI raises $63 million betting that watching how employees actually work is the missing layer of…

What happened

I'll read the post-writing rules and the source summary so the paragraphs stay factual and match the required structure.Skan AI said Wednesday that it has closed a 63 million dollar Series C co-led by Cathay Innovation and Dell Technologies Capital. The company is not pitching another generic enterprise chatbot. It is pitching a layer it calls a context graph of work, built by observing how employees actually perform their jobs across enterprise software rather than how those jobs are written down. VentureBeat carried the announcement. The facts that matter are the 63 million, the Series C stage, the two co-leads, and the product claim that watching real work is the missing layer of enterprise AI.

A context graph of work is a map of practice, not a copy of a process binder. If Skan AI is observing employees across enterprise software, the raw signal is the sequence of screens, fields, tickets, and handoffs a person actually touches to finish a case. Those sequences compress into nodes and edges: a system or screen as a node, a transition as an edge, a repeated path as a weighted route, a detour as a branch that the official SOP never named. The graph is the structure an agent can query. What did this role do after the ERP rejected the invoice. Which CRM field is written only after the spreadsheet is updated. Which exception path is the real path. That mechanic is closer to a runtime trace of human-computer work than to a static diagram drawn in a workshop.

The technical detail

Enterprise software is not one application. A claims analyst, a treasury ops person, or a warehouse planner lives in a stack of systems that do not share a session. The missing context for an AI agent is rarely the schema of any one API. It is the unwritten choreography between those APIs: the order, the retries, the shadow spreadsheet, the approval that happens in chat, the field that is ignored because it is always wrong. A graph built from observation is an attempt to make that choreography first-class data. For an engineer wiring an agent to SAP, ServiceNow, or Salesforce, the graph is the memory the agent does not get from an OpenAPI spec. Without it, the agent calls the documented happy path and then stalls on the case that real staff already know how to route around.

Why it matters for builders

That is why the raise is a bet that observation is the missing layer of enterprise AI, which is exactly how the company framed it. Builders shipping internal copilots already know the failure mode. The model can write SQL and click a UI. It cannot tell a real exception from a noisy one unless someone encodes the exception. Encoding by interview does not scale and decays the week the process owner leaves. Encoding by watching the work is the alternative Skan AI is selling. For platform teams, the implication is architectural. If the context graph sits under the agents, then retrieval, evaluation, and guardrails all attach to observed work rather than to a prompt pack. If it does not, every agent remains a thin client on top of tribal knowledge.

The market this money is entering is not empty. Process mining already reconstructs event logs from systems of record. Task mining already records desktop activity. RPA already replays the clicks. What Skan AI is doing is renaming and re-aiming that stack at agents: not find waste, not automate the keystroke, but give a model a map of how the business actually runs. Cathay Innovation co-leading Series C is a continuity signal on the same thesis. Dell Technologies Capital co-leading is a distribution and infrastructure signal. A corporate venture arm does not need to be a customer to matter. It tells you the thesis is being underwritten by people who sit next to the enterprise account. The competitive question is whether a context graph of work is a new artifact or a rebrand of task mining with an agent API. Buyers will decide that by whether the graph is queryable, versioned, and usable as a runtime dependency, or whether it is another dashboard of discovered processes.

Market and competitive context

What to watch next is not another press line. Watch whether Skan AI publishes a concrete contract for the graph: how traces are captured across enterprise software, what an agent is allowed to read, and how a change in observed practice updates the map. Watch whether Dell Technologies Capital stays a check or becomes a channel into the same estates where employees already work in Dell-adjacent stacks. Watch whether Cathay Innovation's continued lead means the company is still a point solution or is being pushed to become the default context layer other vendors plug into. For an engineering org evaluating this class of product, the practical test is narrow. Pick one messy cross-system workflow. Ask whether the vendor can show the observed path, the official path, and the delta, then whether an agent can act on that delta without a human rewriting the prompt.

What to watch next

The risks sit in the observation itself. Watching how employees work is a surveillance product until the capture surface, retention, and access control are specified. A graph of work that includes every keystroke and screen is a liability graph. A graph that is too coarse is a slide. There is also a validity problem. Observed practice includes workarounds that exist because the system is broken. Training an agent on those traces can automate the workaround and lock in the defect. Related prior art is not only process mining. It is also session replay, digital adoption platforms, and the older BPM suite that claimed to model work and then modeled the workshop. Skan AI is raising 63 million dollars on the claim that this time the model is built from the job as performed. The open question is whether that graph stays a research artifact for consultants or becomes the runtime context layer enterprise agents actually call.