AI Can't Transform Work It Can't See

01 /The Concept

In a well-run enterprise, the processes described in documentation, the processes demonstrated in training, and the processes employees actually execute are three different things. The documented version was written during an implementation project. The training version was designed for an idealized user. The real version was adapted, compressed, workarounded, and optimized across a thousand human decisions that nobody recorded.

That's the problem Skan AI was built to solve. Its desktop technology observes how employees actually use enterprise software across every application – mainframes, legacy desktop tools, VDI environments, web apps, office suites – without requiring any integration project. From those observations, it builds a Context Graph of Work: a live map of how operations actually run, including the tools, the sequences, and the logic behind each step.

The graph is the foundation for three products. Blueprint models the current operating state, identifies automation opportunities, and produces a sequenced AI implementation roadmap with projected ROI per recommendation. Process Intelligence surfaces bottlenecks and ranks automation candidates by complexity and impact. Agents deploys AI systems grounded in actual workflow, not a process document's approximation of it.

Privacy architecture is central here. Raw screen content stays inside the customer's environment. Only anonymized metadata reaches Skan's cloud. Processing runs on Nvidia AI Enterprise hardware inside the customer's perimeter. The certification stack covers ISO 27001, ISO 42001, SOC 2, GDPR, CCPA, and HIPAA – the requirement set for financial services and healthcare, which make up the core of its customer base.

Seven in ten of the largest US banks, one in four Fortune 50 companies, and three in five of the top US insurers are customers. Revenue grew more than 300% year over year for two consecutive years. Net dollar retention sits around 150%.

Skan raised a $63M Series C led by Cathay Innovation and Dell Technologies Capital, with Citi Ventures, Bloomberg Beta, State Farm Ventures, and Wipro Ventures participating. Total funding is approximately $120M. Co-founders Avinash Misra and Manish Garg have been building this thesis since the company's founding.

02 /Why It Matters

Gartner's estimate is that only 8% of enterprises currently have AI agents in production. An MIT study found roughly 95% of enterprise generative AI pilots failed to show measurable returns. Both data points locate the same failure mode: AI models are capable, but the context they need to work reliably inside a specific organization – the operational specifics, the terminology, the legacy system interactions, the undocumented workarounds – doesn't exist in any form those models can consume.

Skan's numbers from a single bank illustrate the scale of what's missing. Across 1,500 finance staff, the platform observed 11.2 million context switches – the moments when an employee moves from one application to another to complete a task. Mapping those transitions surfaced $37M in operational friction. That friction was invisible before, not because nobody was looking, but because it existed at a resolution that organizational process mapping had never been designed to capture.

The competitive frame is worth examining carefully. Process mining vendors like Celonis extract work signals from system event logs: when a record was created, which user updated it, how long it sat in each stage. RPA vendors like UiPath record screen interactions to build automation scripts. Skan's differentiation is semantic rather than mechanical – it builds models of what decisions are being made and why, not just which buttons were pressed and when. That distinction determines whether the resulting AI can handle exceptions, not just the nominal path.

Net dollar retention at approximately 150% is the metric that matters most for this business. Enterprise platform expansions of that magnitude mean customers aren't staying for convenience – they're expanding scope as the context graph becomes more complete and the agents become demonstrably more reliable. At one bank's AML operation cited in the funding materials, 60% of cases now run end-to-end through Skan's AI agents. That is not a pilot metric.

03 /Opportunities

Skan's core insight – that AI transformation requires an observational grounding layer, not just a better model at the application level – generalizes to every professional domain where work is performed through a specific, institutional combination of tools, terminology, and learned sequence.

The most tractable adjacent application is professional services: law firms, consulting firms, accounting practices. These organizations perform knowledge work at high complexity and high specificity – the same conditions that make enterprise AI fail without a context layer. The procurement cycle is shorter than enterprise: a mid-size law firm makes a tool decision in weeks, not months, and the decision often rests with one or two partners. A context graph for how associates research and draft documents could ground AI at a level of specificity that current legal AI tools, trained on public legal text, consistently fail to reach.

Engineering Intelligence is already a Skan product, but the observation goes further than productivity metrics. A context graph built from how a specific engineering team navigates and modifies a specific codebase – which files are always touched together, which test patterns precede which kinds of changes, where deploys consistently bottleneck – would make AI coding assistance significantly more accurate for that team than any approach that reads code without reading the work of writing it.

The pattern for builders: in any domain where knowledge work is both complex and highly specific to an organization, the team that maps the actual work before attempting to automate it will operate with a higher accuracy ceiling than competitors deploying generic models. The observation infrastructure is the defensible layer. The direction worth pursuing: identify a professional vertical where the gap between documented process and actual practice is largest, and where that gap is currently causing AI deployments to fail visibly.