Skan AI, a seven‑year‑old startup that says it builds a "context graph of work" by watching how employees actually perform tasks across enterprise software, announced on Wednesday that it raised $63 million in a Series C round co‑led by Cathay Innovation and Dell Technologies Capital. Citi Ventures, Bloomberg Beta, State Farm Ventures, and Wipro Ventures also participated, bringing the company’s total funding to about $120 million.
Alongside the financing, Skan launched general availability for two new products, Skan AI Blueprint and Skan AI Agents. Together with the company’s existing Skan AI Intelligence offering, these elements form what Skan describes as a full platform for discovering, modeling and automating enterprise workflows.
Context: enterprise AI pilots often fail
The announcement comes amid growing frustration with enterprise AI deployments. According to research Skan cites from Gartner, only 8% of enterprises have AI agents in production and 95% of early implementations will require a redesign. An MIT report covered last year reached similar conclusions about a large share of generative AI pilots failing to produce measurable returns.
Avinash Misra, Skan’s co‑founder and CEO, argues the problem is not the models themselves but their lack of an accurate view of how businesses operate. He says the common grounding strategy—feeding agents process docs, SOPs and system logs—rests on a fiction: documented processes often differ materially from how work actually happens, and agents fail in that gap.
The approach: screen‑level observation and abstraction
Skan’s solution is to observe work at the source. The company deploys software on employee desktops that continuously watches how work flows across applications—spreadsheets, CRM, email clients, legacy mainframes—and abstracts those observations into a living model of the underlying business process.
Skan contends that backend logs capture committed states (completed transactions) but miss the messy human work that happens between those states. The company argues the screen is where human agency, the entire application landscape, and the relevant data converge, and that decades of UI design have buried process knowledge in the space between a worker’s eyes and their monitor. The core technical challenge, Misra says, is not the act of observation but the abstraction: extracting intent and statefully classifying actions at enterprise scale.
From these abstractions Skan builds a context model that AI can reason over and act upon—this model underpins the company’s agent product and the broader platform.
Privacy, surveillance concerns and safeguards
Continuous screen observation raises privacy and workplace surveillance concerns. Reuters reported in June that Meta scaled back an internal tool tracking employee mouse clicks after workers raised issues, illustrating how sensitive the balance can be.
Misra says these worries were anticipated and influenced Skan’s architecture: the system aggregates rather than individualizes behavior, surfacing statistical patterns across hundreds of workers performing the same process rather than tracking any single person. Organizations scope what the technology can observe via opt‑in rules that specify applications and URLs, and Skan says the data it produces does not leave the customer firewall. A three‑tier architecture transmits only anonymized metadata to the cloud. Misra points to deployments approved by European works councils and to security reviews that cleared Skan where many AI tools cannot operate.
He also acknowledges that aggregated telemetry can reveal underperforming teams and that some customers have used insights to reduce headcount in certain processes.
Claimed customer value and concrete results
Skan claims more than $500 million in cumulative customer value to date. Misra clarified that this figure reflects cumulative, customer‑identified and quantified savings opportunities across the customer base—an envelope of expected savings rather than a bank balance of realized cash.
More tangible results from individual deployments include: at one large U.S. bank Skan observed 11.2 million context switches across 1,500 finance professionals and identified $37 million in operational friction. Implementing agent‑executable context reduced cost per transaction by 32%, increased throughput by 41%, and produced $18 million in annualized savings. In an anti‑money‑laundering operation at another bank Skan reports that 60% of cases are now handled by AI agents and that agent accuracy can exceed human accuracy. Among insurers Skan typically reports roughly 25% productivity uplift in core claims processes; one insurer doubled case volume over the past year without adding staff.
Publicly referenceable customers include Unum and Mitie. Skan declined to disclose revenue but said it grew more than 300% year‑over‑year for a second consecutive year, has net dollar retention around 150%, and counts seven of the ten largest U.S. banks and a quarter of the Fortune 50 among its customers.
Learning from imperfect work
Skan’s thesis depends on observing real work, which includes mistakes and inefficiencies. Misra compares the approach to large language models: just as those models learned semantic structure from a mixture of good and bad text, Skan’s models ingest many end‑to‑end process executions and learn the distribution of paths—efficient, slow, compliant or non‑compliant—without assuming a single path is optimal. Organizations then constrain the model by the dimensions they care about (for example, compliance), and the model returns paths that satisfy those constraints.
Misra contrasts Skan’s approach with record‑and‑play automation (common among some RPA vendors), arguing Skan builds a model that understands work and then constrains it, rather than merely replaying observed actions.
He gives a concrete example: at one large bank Skan’s telemetry continuously compares live case execution to a 600‑page controls inventory, with agents that trigger alerts when required compliance steps are missed—turning a static document into a real‑time enforcement layer.
Competitive positioning and strategy
Skan positions itself between process mining, RPA and platform vendors. Process mining sees only backend logs; RPA often replays tasks without understanding them; platform giants (ServiceNow, Salesforce, Microsoft) build agents that are usually constrained to their own environments. Misra argues that work spans systems and the context layer must reflect that breadth.
The company emphasized a partnership with Nvidia: Skan runs on Nvidia AI Enterprise and NIM microservices, and there is growing demand for private appliances that can observe work, hold the context model, and execute agents entirely within a customer’s infrastructure. Investors have predicted enterprises will spend more on AI in 2026 but through fewer vendors, a consolidation that could favor a vendor offering discovery, intelligence and agents as a closed loop.
Misra’s summation is blunt: "You cannot retrieve context that you do not capture." He argues the battleground for enterprise differentiation will shift from the smartest model to knowing precisely how a company works—context that the company itself protects.
Conclusion
With $63 million in new capital and a product stack that combines discovery, modeling and agent execution, Skan is betting that mapping the hidden, screen‑level reality of work will become foundational infrastructure for enterprise AI. The approach promises measurable operational gains for some customers, but it also raises enduring questions about privacy, governance and how organizations choose to act on the telemetry they collect.
