Most companies are trying to teach AI to work the way they *wish* people worked, not the way people actually do.
The Summary
- Skan AI raised $63M Series C to build a "context graph of work" by observing how employees actually perform jobs across enterprise software
- Only 8% of enterprises have AI agents in production, and 95% of early implementations will need complete redesigns (per Gartner)
- The company argues enterprise AI fails because models are trained on process documentation that doesn't match how work actually gets done
- Skan's new Blueprint and Agents products aim to discover, model, and automate workflows based on observed reality, not fictional SOPs
The Signal
The dirty secret of enterprise AI: every company has two operating systems. There's the one in the official process docs, the Confluence pages no one reads, the SOPs written by consultants who left three years ago. Then there's the actual system, where Sarah in accounting has a spreadsheet workaround, where the CRM integration breaks so people copy-paste into Slack, where the "approved workflow" takes 14 clicks but everyone knows the 3-click shortcut.
Skan AI's thesis is that you can't automate the second system by studying the first one. Their software watches employees work, across every application they touch, building what they call a context graph of actual workflows. Not the org chart version. The real version.
"Everyone is obsessed with building a better driver. We think the bigger opportunity is building a better navigation system."
This matters because the $120 million Skan has now raised is a bet that the entire enterprise AI stack has been built on the wrong foundation. The models are fine, CEO Avinash Misra argues. GPT-4 can reason. Claude can follow instructions. What they can't do is operate in businesses they don't understand.
The numbers back up the pain. Gartner data shows 8% of enterprises have AI agents in production. Not "working well." Just "in production." And 95% of the early attempts will need complete redesigns. That tracks with MIT's 2024 finding that 95% of generative AI pilots failed to deliver measurable returns.
Key failure modes:
- AI trained on official process docs that no one actually follows
- Agents that can't handle the informal shortcuts and workarounds that make work actually flow
- No map of how information really moves between systems and people
Skan's platform has three layers now. Intelligence observes and maps workflows. Blueprint models them. Agents automate them. The sequence matters. You can't automate what you haven't accurately modeled, and you can't model what you haven't observed in the wild.
The customer base suggests this isn't just theory. Skan works with Citi, State Farm, and Wipro, companies where the gap between "how we say we work" and "how we actually work" is measured in billions of process inefficiency. Dell Technologies Capital and Citi Ventures both put money in this round, which means the people buying enterprise software and the people funding it both think observation matters more than documentation.
The Implication
If Skan is right, every company trying to deploy AI agents needs to add a step before prompt engineering: reality mapping. Before you automate the workflow, watch the workflow. Before you build the agent, build the map.
For builders in the agent economy, this is the infrastructural layer most are skipping. You can have the best LLM, the cleanest API architecture, the slickest UI. But if your agent operates on a fictional understanding of how work flows, it will fail in production every time. The companies that win will be the ones that observe first, model second, automate third.