In this exclusive interview with Swapnil Bhartiya at TFiR, Binny Gill, CEO at Kognitos, explains how neurosymbolic AI and natural language automation are transforming enterprise operations — making business process automation accessible, auditable, and truly intelligent for the first time.

The Problem With Layering AI on Top of Old Automation

Most enterprise automation platforms today take a familiar shortcut: wrap a natural language interface around an existing RPA or workflow tool and call it AI-powered. Binny Gill argues this approach is fundamentally broken — not because the tools are bad, but because generative AI is the wrong engine for execution. Making a generative model smarter doesn’t make it more predictable; it makes it more creative, and in business automation, creativity at runtime is a liability.

Kognitos takes a structurally different approach. Instead of using large language models (LLMs) to plan and execute at the same time, the platform separates the two functions entirely. Planning — the creative, generative work of turning a business requirement into a step-by-step process — is handled by a neural AI layer. Execution is handled by a symbolic engine built in Rust that runs English-language SOPs deterministically, without LLMs, and therefore without hallucination.

Q: What made this approach possible now, and how much involvement does it take from a non-developer business user?

Binny Gill: “Five years ago, this was pre-ChatGPT and others. I wanted to figure out, is this the right time when we can democratize the art of telling a computer what to do — from just developers to everybody in the business? For the last five decades, what was happening was the business side would get an idea. Somebody sitting in finance and accounting would say, ‘Next quarter, I would like to do it this way. This is how I pay invoices. This is how I onboard new vendors.’ But then they have to give it to IT. A developer translates it into machine language, whatever the machine understands. My inspiration actually came from Alexa and my kids talking to Alexa in English. They’re telling a computer what they want and it does it for them. Why does business have to be any different?”

Q: How does Kognitos differ from companies that just layer AI on top of traditional automation tools?

Binny Gill: “The approach has to be different. If you put in a generative AI that is creative, you can’t make it smarter and expect it to behave the same way. Sometimes a smarter model will become more creative in a bad way, and then it won’t do what you really wanted it to do. You don’t even have to be a prompt engineer — I feel that is just a temporary thing. You just chat with our system and it is going to write your SOP. English is now the new coding language, but the way it writes the English SOP, it’s a very detailed English specification, but it’s fluent English. Anybody can read and understand it.”

Neurosymbolic Architecture: Separating Planning from Execution

The term “neurosymbolic AI” has gained traction in research circles, but Kognitos has built a commercial implementation of the concept specifically for enterprise process automation. The architecture rests on a clear division: the neural layer generates the plan, the symbolic layer executes it. The two never mix at runtime — and that separation is what makes deterministic behavior possible at scale.

Gill draws the analogy to industrial assembly lines: the machine runs continuously and predictably, and humans are only called in when the machine encounters something it cannot handle. In the Kognitos model, that “machine” is the symbolic execution engine. When it hits an edge case — an invoice format it hasn’t seen, a partial payment, a currency translation discrepancy — it doesn’t guess. It stops, escalates to what Kognitos calls the “resolution agent,” which pulls in both a frontier LLM and a human subject matter expert (SME) to resolve the ambiguity. The human’s decision is then encoded as symbolic logic and patched back into the running process, so the same edge case never requires human intervention again.

Q: How do you define neurosymbolic, and why does the separation of planning and execution matter?

Binny Gill: “Neurosymbolic means that when I’m executing, I will be executing with a symbolic engine, but when I’m coming up with a plan, it’s a neural engine. Normally when you’re talking with an LLM, it’s both doing planning and execution at the same time. It’ll plan the next step, then it’ll call. Then again plan, then call. We believe that planning you separate out from execution. Planning can be highly creative and therefore risky. And sometimes it’ll come up with a plan that you don’t want to approve. But once I review it, that must be symbolically executed. It must be executed as if it was code.”

Q: How does this architecture solve the hallucination problem in practice?

Binny Gill: “Our system is deployed in finance and accounting — mission critical — where one extra zero will just bring down the whole company. The AI machine is running deterministic code. It runs the same way. You give it the same input, it’ll give you the same output a hundred times. A prompt does not give you the same answer a hundred times. But in our system, when you’re looking at an SOP in English, run it 100 times, it gives you the same result a hundred times. Anytime there’s a variant, it just cannot do it because it’s not generative AI, it’s symbolic AI. So it pauses and calls the generative AI into the picture. We call it the resolution agent.”

The “Time Machine” Engine and Tribal Knowledge Capture

One of the most technically distinctive components of the Kognitos platform is what Gill calls the “time machine” — a full execution recorder built into the symbolic interpreter. Unlike traditional programming languages or RPA tools, which typically log only errors or final states, the Kognitos engine records every step of every transaction. This has two consequences: first, it enables the resolution agent to understand exactly what went wrong and why, without needing to reproduce the error. Second, it generates a complete, queryable audit trail that satisfies the compliance requirements of regulated industries.

The platform also addresses one of the most persistent challenges in enterprise knowledge management: tribal knowledge. When experienced employees leave an organization, the informal rules and judgment calls they carry in their heads leave with them. Kognitos makes that knowledge explicit and persistent. Each time a human SME resolves an edge case, the system prompts them to decide whether that resolution should become a standing best practice — and if so, it writes a knowledge-based article, appended to the SOP, that captures the rule in plain English.

Q: You mentioned the time machine — can you explain what it is and why you built it?

Binny Gill: “We built it as a time machine in order for AI to be able to understand when things are going wrong, why are they going wrong. But as a side effect, what we get is detailed audit. You can go and ask any question for one particular transaction. What was the value in step two? Why did this get — if-then-else — what happened? Because it’s mapped to business logic deterministically and everything is recorded. It is able to give you a better audit than even earlier RPA used to do. Programming languages don’t give you audit. They just give you the final crashing state. We are recording everything. It’s the most detailed audit that you can get.”

Q: How does this connect to the concept of tribal knowledge?

Binny Gill: “We are collecting tribal knowledge from people’s heads. It’s transparent — they don’t know that it’s growing. The standard process is documented, the tribal knowledge is documented, and this is the layer of indirection between intelligence and what the business wants. AI does not belong to any tribe. You can’t assume that AI will do what you think your employees would have done. It doesn’t know which company it’s working for and how the C-suite of that company really wants to act.”

Real-World Deployment: Fortune 500 International Payment Reconciliation

Kognitos has deployed its platform across finance and accounting, HR, and legal workflows. Gill’s most detailed example involves a Fortune 500 company that had previously attempted to automate its international payment reconciliation process using a major RPA vendor — and failed. The process involves matching daily bank wire transfer reports against SAP DSO (Days Sales Outstanding) data, across a mailbox receiving roughly 1,000 emails per day, with payment confirmations arriving as emails, PDFs, screenshots, and images with typos, partial payments, and multi-invoice bundles.

RPA broke constantly on the variation. The eight-person manual processing team didn’t shrink — it was replaced by four RPA developers earning significantly more, with no net efficiency gain. Kognitos implemented a process-driven approach: the system extracts invoice numbers and amounts from any document format, performs fuzzy matching against the SAP DSO report, and handles perfect matches automatically. Edge cases escalate to the resolution agent. Over time, the system learned that a sub-dollar discrepancy is acceptable, that yesterday’s exchange rate is sometimes used, and dozens of other business-specific rules — encoded directly from SME feedback. Automation rates moved from 80% on day one to 97% over time, with the remaining 3% representing genuine exceptions requiring human judgment.

Q: Can you walk us through a specific Kognitos deployment?

Binny Gill: “They’re a Fortune 500 company — all their international payments. Before Kognitos was put in, they had tried to do it with one of the RPA vendors. People wire the money to the company and it goes to one of three banks. Every day the bank report comes in — this wire transfer is obtained from this particular name, and in that memo is what invoice is being paid. This has to be matched with the DSO report from SAP every day. At the high level it seems very simple. You apply RPA to it and it keeps on breaking every day. Because people wire the money — sometimes they put a screenshot, sometimes a PDF, sometimes there’s a typo, sometimes a partial payment, sometimes they club two invoices and pay in one. All sorts of variations. RPA would constantly break.”

Binny Gill: “First time we did it, 80/20 rule — 80% would go through. Remaining 20%, it would raise a hand, go to the same person who used to do it manually and say, ‘Hey, I couldn’t do it.’ Manual person says, ‘Okay, there’s only less than a dollar difference. In this case, we normally accept it.’ Now, AI cannot guess that business is okay with a dollar difference. And now it learns — says, ‘Okay, should I learn this thing?’ And say yes. So now it has learned that tribal knowledge. Next time it doesn’t bother. So over time, our system has learned this thing. 80% has now become 97%.”

The Future Enterprise: Philosophy Over Automation

Gill’s long-term vision for Kognitos goes beyond process efficiency. In his view, automation will eventually be so commoditized that it ceases to be a strategic differentiator. What enterprises will compete on instead is their documented philosophy — the values, priorities, and approaches that define how they want to serve customers — translated into AI-executable language. CEOs will have philosophical conversations with AI systems that curate and encode those values into every operational process. The company with the clearest, most thoroughly documented philosophy will win.

Q: If your vision plays out, what does the enterprise look like in a few years?

Binny Gill: “I think we will stop talking about automation, or we’ll take automation for granted. The difference between one enterprise and the other would be about philosophy — how do we want to approach solving the problems for our customers? Just like every restaurant has a different cuisine and a different way of serving customers, but cooking is not a differentiator. It’s like everybody can cook. But how you present, what is the ambiance — that is what enterprises will differentiate on. CEOs would be having a philosophical conversation with AI and saying, ‘This is what I care about.’ CEO of Patagonia would say, ‘I care about environment.’ CEO of McDonald’s would say something different. So automation, SaaS, software — these would get commoditized.”