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

What happened

Skan AI has raised 63 million dollars in Series C funding, a round co-led by Cathay Innovation and Dell Technologies Capital, the company announced Wednesday. The product it is selling is not another foundation model and not another chat interface on top of a system of record. Skan AI builds what it calls a context graph of work by observing how employees actually perform their jobs across enterprise software. The title of the raise is the thesis: the missing layer of enterprise AI is a living map of how work is done, not how it is documented. That is a funding event with a specific architectural claim attached to it.

A context graph of work is not an org chart and it is not a process diagram produced in a workshop. The method is observational. Skan AI watches people move through the enterprise applications they already use, then assembles those traces into a graph of real sequences, handoffs, exceptions, and the informal paths that never appear in a standard operating procedure. In that construction, nodes are the systems, screens, records, and decision points a person actually touches. Edges are the transitions they take, including the ones that skip a required field, bounce to a second system, or wait on a Slack message that no BPMN file will ever mention. That is a different data substrate than CRM rows, ticket text, or the prompt logs of a copilot. It is behavioral telemetry of work, structured so a machine can reason about the job as it is performed rather than as it was designed.

The technical detail

Engineers shipping agents into the enterprise already know the failure mode. A model can propose a plausible next action and still be wrong about which system to open, which field is the source of truth, or which exception path a human takes when the happy path breaks. Fine-tuning on wikis and tickets does not capture click-level procedure. Retrieval over policy documents does not tell an agent that invoice exceptions are resolved in a different application than the one that created the invoice. A context graph of work is an attempt to give agents the same kind of grounded state that process mining promised for robotic process automation, but aimed at generative systems that need to decide, not just replay a recorded macro. Anyone wiring a copilot into Salesforce, SAP, ServiceNow, or a private tangle of internal tools will recognize the gap. The model is not the bottleneck. The missing object is a current, queryable description of how the work actually moves.

Why it matters for builders

The raise sits in a crowded neighborhood. Process mining and task mining vendors have spent years reconstructing process graphs from event logs and desktop traces. RPA platforms automated the happy path and then struggled with the long tail of exceptions that make up most of real work. A newer wave of enterprise AI vendors is selling agents that claim to do the work rather than describe it. Skan AI is positioning the context graph as the layer those agents lack. Cathay Innovation and Dell Technologies Capital co-leading a Series C is not a casual pairing. Cathay is a global growth firm with a long enterprise-software book. Dell Technologies Capital sits next to one of the largest channels into corporate IT. That combination says the company is selling a platform story into existing stacks, not a consumer-style demo of an agent that looks competent in a recorded walkthrough.

What to watch next is whether the graph becomes an integration surface. If Skan AI can expose the observed work graph to the systems of record and to the agents other vendors are deploying, it becomes infrastructure. If the graph stays locked inside Skan's own product, it is another process-intelligence suite competing on dashboards and transformation slideware. Engineers evaluating this class of tool should ask three concrete questions. What is observed, at what granularity, and with whose consent. How is the graph updated when a process changes, which they always do, often without a ticket. And can an agent query the graph at runtime, or is the output a report that a consultant reads once a quarter. The 63 million dollars will be spent answering those questions in customer environments, not in a research blog.

Market and competitive context

The open risk is the same one that has followed every workplace-observation product. Watching how employees work across enterprise software is a privacy, labor, and security problem before it is an AI problem. Application-level or screen-level telemetry can leak credentials, customer data, and performance signals that get used for surveillance rather than process design. There is also a representation problem. A graph built from observed clicks will encode the workarounds, tribal knowledge, and sometimes the policy violations that keep a process running. An agent that imitates that graph will imitate the shortcuts. Skan AI has not, in the announcement summarized here, published the technical boundary between observation and recording, or the retention and access model for the graph. Those details will decide whether this is a missing layer or a new liability sitting on the critical path of how work is done.

What to watch next

The prior art is not mysterious. Celonis and the rest of the process-mining cohort reconstruct process variants from system event logs. Task-mining tools sit on the desktop and reconstruct the human path across applications. Digital-adoption platforms overlay guidance on the same screens. What Skan AI is claiming is that those traces, assembled as a context graph, are the right memory for enterprise AI rather than a quarterly report for a transformation office. That claim will be tested by whether the Series C, co-led by Cathay Innovation and Dell Technologies Capital, buys distribution into the environments where work already happens, and whether customers will let observation software sit close enough to employees to build a graph that is actually current. A context graph that is six months stale is just another stale process map with a newer name.