Key takeaways
- On August 12, 2026, Skan AI announced a $63 million Series C co-led by Cathay Innovation and Dell Technologies Capital, to ground AI agents in the context of real work.
- Skan AI's platform has processed 25 billion work signals; it serves 7 of the 10 largest US banks and a quarter of the Fortune 50, with more than $500 million in measured customer value according to the company.
- At a top US bank, observing 1,500 finance professionals surfaced $37 million in operational friction; cost per transaction fell 32% and annualized savings reached $18 million.
- Market context: according to MIT's GenAI Divide study, 95% of enterprise generative AI pilots deliver no measurable P&L impact.
What Skan AI announced on August 12
On August 12, 2026, Skan AI announced a $63 million Series C co-led by Cathay Innovation and Dell Technologies Capital, with participation from Citi Ventures, Bloomberg Beta, State Farm Ventures and Wipro Ventures, according to the official press release. The round comes with the general availability of a three-part platform: Skan AI Blueprint to discover and prioritize AI opportunities, Skan AI Intelligence to manage workflows, and Skan AI Agents for autonomous execution grounded in business context.
The company's thesis fits in one image, coined by co-founder and CEO Avinash Misra: everyone is obsessed with building a better car; the bigger opportunity is building a better navigation system. In other words, AI models are already powerful; what agents lack is the map of the terrain, a precise knowledge of how a company actually executes its processes.
To produce that map, Skan AI's technology observes how employees and systems carry out the work, then turns those observations into context that agents can use. The company says it has processed 25 billion work signals, reports 300% year-over-year growth and 150% net dollar retention, and counts 7 of the 10 largest US banks plus a quarter of the Fortune 50 among its customers, notes Pulse2. The platform runs on NVIDIA AI Enterprise and NIM microservices, with a research partnership with the University of Missouri.
Why is context the real bottleneck for AI agents?
Because the failure numbers are massive and documented. MIT's study "The GenAI Divide: State of AI in Business 2025", published in the summer of 2025, estimated that 95% of enterprise generative AI pilots deliver no measurable P&L impact, despite $30 to $40 billion invested, as Forbes reported. MIT's diagnosis did not blame model quality; it blamed the lack of integration into real workflows. A year later, Skan AI's round is an investor response to that precise diagnosis: the problem is not the engine, it is knowledge of the terrain.
The case study highlighted in the press release illustrates the method. At a top US bank, Skan AI observed 11.2 million context switches, those jumps from one screen or tool to another, across the daily work of 1,500 finance professionals. The analysis identified $37 million in operational friction. After redesigning the affected processes, the bank reports a 32% drop in cost per transaction, a 41% increase in throughput and $18 million in annualized savings. The raw material for those gains was not a better model; it was close observation of existing work.
The composition of the round tells the same story. Dell Technologies Capital, Citi Ventures, State Farm Ventures, Bloomberg Beta, Wipro Ventures: funds backed by an industrial group, a bank, an insurer, an IT services firm. These are field buyers, not just financiers. Skan AI also claims average operational savings of 30 to 40% for its customers; that figure comes from the vendor and deserves the usual caution, but the market's direction is clear: after models, then agents, the money is moving to the context layer.
What does this change for an SME?
The transferable lesson fits in one sentence: before deploying an AI agent, you need to know how the work actually gets done, not how the procedure says it gets done. In a company of 10 to 100 employees, the gap between the two is often considerable: the Excel export nobody documented, the re-keying between two tools that do not talk to each other, the informal check an experienced employee performs from memory. An AI agent wired to the official procedure automates a fiction; that is one reason so many pilots produce nothing.
An SME does not need a process intelligence platform sized for Fortune 50 banks. It needs the same discipline, at its own scale. The table below sums up the three ways to capture the context of real work, from simplest to most instrumented.
| Method | Entry cost | What it reveals | Who it is for |
|---|---|---|---|
| Manual mapping: interviews and desk-side observation | A few days of work | Gaps between official procedure and real work, informal checks | SMEs of 10 to 100 employees, before any automation project |
| Analyzing logs from existing tools: CRM, email, ERP, n8n | Low, the data already exists | Real volumes, lead times, measurable bottlenecks | Companies whose tools already trace activity |
| Dedicated process intelligence: Skan AI, Celonis and peers | Enterprise project, custom pricing | Continuous, exhaustive observation of on-screen work | Large accounts, thousands of users, high-volume processes |
Judgment cuts plainly here: if your processes fit on a whiteboard after a morning of interviews, an automated observation tool is oversized, and the budget is better spent on the redesign itself. Row three of the table is not for you, and that is good news: at your scale, the missing layer costs a few days of rigor, not an enterprise subscription.
That is the reflex my projects start with. Before writing a single line of the Emma CRM for 3018, France's national hotline against cyberbullying that I co-built for the e-Enfance association, the work was to understand how 12 counselors and managers actually handle cases across 5 channels; the tool was drawn on that reality, not on an org chart. Same logic for the automated market watch I built for the Ermitage cheese dairy: first follow the real path the information takes, then automate it. An AI agent that ignores how work actually gets done automates the theoretical version of the job; it is the software equivalent of the procedures binder nobody opens.
The question to ask before any AI agent project is therefore not "which model should we pick?" but "who, in our company, can describe what really happens between the order and the invoice?". If nobody can answer, that is the first project. And it costs a lot less than $63 million.
Frequently asked questions
What is process intelligence?
Process intelligence means observing, through digital traces or on-screen work capture, how processes actually run inside an organization, then comparing that reality with the official procedure to target improvements. Skan AI, which raised $63 million on August 12, 2026, applies this approach to give AI agents business context.
Why do 95% of enterprise AI pilots fail?
According to MIT's study "The GenAI Divide: State of AI in Business 2025", 95% of generative AI pilots deliver no measurable P&L impact, mostly because they are not integrated into real workflows, not because of model quality. The study also found that tools from external vendors succeed about twice as often as internal builds.
Does an SME need a process intelligence tool?
Rarely. These platforms are sized for large-enterprise volumes. An SME gets most of the benefit from manually mapping its processes, a few days of interviews and observation, plus analyzing the logs of its existing tools, before any AI agent or automation project.