The Real Reason Enterprise AI Agents Keep Failing
Enterprise AI is sitting on a dirty secret. Companies have spent billions standing up generative AI pilots, yet the results are embarrassing. Gartner research cited by Skan AI finds that only 8% of enterprises have AI agents in production, and 95% of early implementations will require a complete redesign — figures that track closely with an MIT study covered by Fortune, which found a similar 95% failure rate for measurable returns from generative AI pilots.
Avinash Misra, co-founder and CEO of Skan AI, has a pointed diagnosis: the models aren't the problem. The problem is that every process document, SOP, and system log fed to those models was never an accurate picture of how work actually gets done.
"Everyone is obsessed with building a better driver. We think the bigger opportunity is building a better navigation system."
It's that conviction — and seven years of building around it — that now has investors writing large checks.
$63M and a Complete Platform
The new $63M Series C, co-led by Cathay Innovation and Dell Technologies Capital, takes total funding to roughly $120M. Citi Ventures, Bloomberg Beta, State Farm Ventures, and Wipro Ventures also participated.
Alongside the raise, Skan is launching two new products — Skan AI Blueprint and Skan AI Agents — which, together with its existing Skan AI Intelligence offering, form an end-to-end platform for discovering, modeling, and automating enterprise workflows.
The "Context Graph of Work" — and Why Logs Aren't Enough
Skan's core product is what the company calls a "context graph of work": a continuously updated model of how employees actually move through their jobs across enterprise software. The mechanism is desktop observation — the platform watches how work flows across applications, from spreadsheets and CRMs to 40-year-old mainframes, and abstracts those observations into a living process model.
This directly challenges the dominant approach used by process mining vendors like Celonis, which reconstruct workflows from backend system logs. Misra's critique is structural:
"All backend data, by definition, is a committed state of work. Work is really what happens between those committed states."
In his framing, system logs only capture the finished transaction — not the decisions, rework loops, exceptions, and institutional habits that produced it. He puts 80% of what AI agents actually need to understand in that invisible gap between systems, and argues the screen is the only place where human agency, application data, and process knowledge converge.
The hard part, he emphasizes, isn't observation itself — it's abstraction. A human watching a colleague's screen can instantly tell whether jumping back to step one means a new case or rework on an old one, because they understand the signature of the work. Teaching a model to make that same judgment, statefully and at enterprise scale, is where Skan believes its seven-year head start lives.
The Surveillance Problem — and How Skan Has Tried to Solve It
Continuously watching employee screens is not a neutral product decision, and Misra says he heard the objection before writing a line of code. When he first pitched the concept to the chief transformation officer at AXA Mexico, her reaction was immediate: "You are dead on arrival. You are observing things that you shouldn't be observing."
That pushback, he says, directly shaped the architecture. Skan aggregates rather than individuates: it surfaces statistical patterns across hundreds of workers performing the same process, not the behavior of any individual. The platform operates on a strict opt-in scoping model — only specific, pre-approved applications and URLs are observed — and data never leaves the enterprise firewall. A three-tier architecture sends only anonymized metadata to the cloud.
The company points to approvals from European works councils — among the world's most privacy-protective labor bodies — as evidence the design holds. That said, Misra acknowledged the technology has led some customers to reduce headcount in certain processes, which means the ethical questions around aggregation are unlikely to go away entirely. Meta's reported rollback of an internal mouse-click tracking tool after employee pushback is a reminder that even AI-forward companies are navigating the same line between operational telemetry and surveillance.
What This Means for Founders and Operators Deploying AI
For startup founders and enterprise product teams building on top of AI agents, Skan's thesis carries an uncomfortable implication: your source data is probably wrong. If agents are being grounded on process documentation written to describe how work should happen rather than how it does happen, failure is built in — and no amount of model tuning will fix a data problem that lives upstream.
What the round confirms is that investors see real enterprise demand for a pre-agent infrastructure layer: a system that audits operational reality before agents are deployed into it. Whether screen-level observation becomes the standard input layer for enterprise AI — or a niche complement to process mining — depends on how the next generation of agent deployments performs in production. For now, the $63M bet is that the navigation system matters as much as the driver.
