How Skan AI Is Building the Context Layer for Enterprise AI With Avinash Misra

• Most enterprise AI pilots succeed and then fail in production for one reason: the curated pilot never contained the exceptions, edge cases, and lived history that real operations run on.

• Skan observes work directly across hundreds of desktops with zero integration, distilling it into a context layer that lets companies collapse human processes into agentic ones.

• Misra's bet on the future: decisions get ceded to machines, while choice, taste, and goal-setting stay human, making 'all future jobs QA jobs.'

Let's dive in. No floaties needed…

Revenge of the Nerds

Avinash Misra, CEO of Skan AI

Avinash Misra co-founded Skan AI in 2018 with Manish Garg, his IIT classmate and childhood friend from Kanpur, India. They'd already built and sold one company together, Endeavour Software, acquired by Genpact in 2015.

Skan's platform observes work across every application, from mainframes to browsers, with zero integration, then reconstructs the real process. It raised a $40M Series B in 2022 led by Dell Technologies Capital, bringing total funding to $54M, and counts seven banks, insurers, and several Fortune 500 companies among its customers (one investor, Citigroup, is also a customer).

What makes Misra worth listening to is his unfashionable argument in the middle of an agent gold rush: the agent is the easy part, and the hard part is the context nobody has bothered to write down. He frames it through Polanyi's Paradox, 'we know more than we can tell,' arguing that the tacit knowledge in human minds is what kills enterprise AI in production. His sequence is inverted: observe how work happens, build the context plane, collapse the process, then lay AI on top.

Tell us a little about Skan AI

Skan is the operational representation layer for the enterprise when it comes to AI. Foundational models are increasingly capable, but most enterprises don't possess the machine-readable representation of how work actually gets done. Critical execution logic still resides in human minds: exceptions, judgment, and informal coordination. So Skan observes this work directly, happening on hundreds of screens, hundreds of people collaborating, and constructs a living model of enterprise execution that carries that context from the deployment to the governance of AI agents.

The analogy I use is that today's AI gives you the reasoning, and Skan gives you the perception and the operational memory layer.

Can you give a concrete example of how a team uses it day to day?

First, let me characterize our customer base: seven banks, insurance companies, large manufacturers. But the point isn't that they're banks. Three things are common across all of them. One, a large number of people. Two, a technology ecosystem built up over time, from archaic mainframes to the most modern applications. Three, they lack deep knowledge of how work actually happens, because they have to change it to optimize it, or build AI that understands it. It's not enough to say, give me the documentation, because that documentation doesn't exist. It's in the lived execution that our customers find the challenge.

Within those organizations, the first port of call is operations: loan servicing, insurance claims, healthcare prior authorization, customer service. Think of an insurer, from first notice of loss to adjudication to payment. The company may think a claim runs 20 minutes across 15 steps, but 90% of the time there's huge deviation. So when you train an AI on what you think the process is, it never seeks the goal the way a human does. And AI doesn't need to follow the human path. Humans have a smaller context window, and the interface was designed to match it: five fields, then next. AI's context window expands exponentially, so the process collapses. We call it business process collapse. What we produce isn't the standard operating procedure, but the agentic operating procedure.

What does implementation look like and how long until the agent is trained?

First, how long to build the context, because agent training comes later. We deploy a non-intrusive technology. Simple example: if you share your screen and I watch you open Excel, go to a CRM, go to email, come back to Excel, in one or two iterations I build a mental model of your process. If you're diligent, you'd notice that 20% of the time, when the claim is greater than $20k, you follow a different path. But you can't do that at scale, 24/7, across thousands of people. So you give that same ability to a model that observes work simultaneously across all desktops.

Now I can follow individual work. This claim was here; certain things were done; five days later it shows up on someone else's desktop, then comes back. That creates what we call the telemetry of work, which is stateful. We distill it into context and collapse it from a standard operating procedure into an agentic operating procedure. This tells AI agents what to do at each step, and if they get stuck, what was done when humans got stuck in similar cases. The context becomes the execution brain, and the agentic layer becomes the arms and legs.

Does the agent have to be connected to a company's entire tech stack?

This is the beauty of it. When you observed my screen, I could have opened Excel, a mainframe, anything, and you still understood. Our technology is completely integration-free. It observes work like two human eyes and makes sense of it from what it sees on the screen.

Sometimes there's more to a process than two eyes can see. Say I'm a specialist remediating customer problems; I'm in their system, but I also know things from documents I've read. So we let organizations feed our models their standard operating procedures, templates, or knowledge bases. The model then matches what it sees against what the person executing knows. Many times we've shown customers their workforce isn't actually deploying all the policies and constraints, whereas the AI can, because it correlates documentation with the lived execution history. But primarily, we don't require any fundamental integration. It's based on what we're seeing on the screen.

How does Skan compare against Claude or ChatGPT for enterprise automation?

The point isn't Claude or any particular AI. Talking about Claude in a business process is nothing different from 20 years ago. You had large language models then too, except they ran on 20 watts of power. You still needed organization of work, policy control, compliance, governance. So the rise of AI at an individual capability level makes our work even more valuable, because we string these pieces of work together into the business process.

Underwriting a claim may well be a model. But when and how that claim flows through the system, between departments, how you report back to the regulator, how you prove a policy is being applied, that is work. There's another point: deterministic execution. AI tools are fundamentally probabilistic, but a bank or insurer can't take the risk of saying 'probabilistically, this is what will happen.' How do you make a probabilistic tool deterministic? By putting context and harness around it, so of the 20 possible outcomes, only the right one is chosen, not the 21st one the model invents.

Skan launched a standardized vocabulary for human and AI collaboration. Can you walk us through it?

It's not just about Skan. Economists speak in a certain language, doctors speak in a certain language. You can't industrialize a field until you standardize its vocabulary. Every engineering discipline develops standards: software has APIs, networks have protocols, databases have schemas. Human-AI collaboration today lacks a common language. It exists as prompts and human engineering, but without shared definitions, every organization reinvents processes, with no common language around delegation, supervision, escalation, and accountability.

If we built those structures for humans, similar structures need to be built for human-AI work. So we're building our own ontology. We construct a business process by observation, then apply our ontology of what's execution, what's communication, what's decisioning. Based on those, we classify how the AI agent will understand the work, automate it, or report back.

Why do pilots work on paper but fail in production?

This question is the reason Skan exists. Let me put it precisely: pilots don't fail. Pilots succeed; production fails. Pilots operate in curated environments that don't contain the exceptions, edge cases, and lived history. So the intelligence isn't the problem. The AI part, that it can adjudicate, create, decide, isn't the problem. Context is the problem.

Most enterprises dramatically underestimate how much operational context exists in a working process. A pilot fails in production for the same reason a new employee struggles in their first few days. Our goal is to use AI to build the context first, then lay AI on top. This wasn't possible five years ago, because observing work at scale wasn't computationally feasible. Now you can look at a screen and, like a human, ascribe intent: a person is onboarding a customer in 15 steps. If AI does it, those steps collapse into one API call; we don't need the interface. But it all comes from how work is done today. Why Skan exists and why production fails is the same answer: context.

What was your main learning from working with Manish that you applied at Skan AI?

The biggest learning is that whether it's AI adoption or, earlier, adoption of mobile devices in the enterprise, it's not a technology problem. Technology adoption is ultimately a human problem. Across two decades, we learned that technically correct answers often lose out to operationally acceptable ones. With agentic AI, the fact that an agent can produce results has value, but the governance, the explainability, being able to prove to a regulator what was done- that matters even more.

People don't resist technology; they resist uncertainty. The one lesson across the board is that new technology is fundamentally a human psychological problem because it introduces uncertainty. If you can manage that uncertainty, adoption becomes easy. That's our lesson: human psychology, even though we're a technology company.

What does the agentic enterprise look like in 10 years?

One can always wiggle out by saying 'who knows,' but that wouldn't be fun. My sense is strong here. One of my best-performing LinkedIn posts was that all future jobs are QA jobs. Whether you're a lawyer, a software engineer, or a content manager, you essentially do quality assurance. And I mean that in the larger sense: not just quality, but taste, goal, choice. There's a documentary, Plug and Pray, on Weizenbaum, who in the '60s wrote ELIZA, a program that just responded in human language. He became a critic of AI because people didn't understand it was just algorithmic. And he said something beautiful: decision is algorithmic, choice is human.

So in the future enterprise, all decisions are made by machines, and all choices, tastes, and goals are made by humans. QA will ultimately be human, because human choice doesn't come from a decision path. It comes from deep emotion about what the human wants to achieve. My second prediction: today, there's a perversity where you pay a consulting firm millions to be told what your own process is. With AI and companies like us, that changes. Organizations will understand their outcomes and how they produce them in complete clarity. That much I can promise you.

Additional Reads

• Avinash Misra, CEO & Co-Founder of Skan.AI (Interview Series) — Unite.AI

• Fireside Chat: Phil Fersht Talks with Avinash Misra — HFS Research

• Why Agentic AI Fails Without First-Party Data — Skan AI blog

Quick Poll

Rate This Edition