[Ep. 090]

Come Disrupt Your Own Industry: How Imubit Put AI in Control of Live Oil Refineries

Listen on:

Gil Cohen trained as a mathematician and engineer. Now he runs Imubit, a company doing the thing most AI startups go out of their way to avoid. Its AI does not sit in the cloud handing out suggestions for a human to take or ignore. It moves real valves inside working oil refineries, nudging temperatures and pressures every minute, day and night.

Stay on that for a second. A neural network is writing straight back into Honeywell and Schneider Electric control systems that are 40 and 50 years old, in plants where a wrong move can hurt someone. Imubit calls it closed-loop optimization. It is about as literal as Physical AI gets in a hard industry. Not a dashboard. Not a copilot. Software with its hands on the equipment.

Imubit has signed seven of the ten biggest US refiners and raised around $100 million. In this conversation Gil gets into how the technology actually works, why his first hiring plan failed, and where he thinks the whole industry is about to be wrong about AI agents on the plant floor.

The conversation

Jay: I want to start by laying out the landscape of AI in legacy industries. In 2024, the Dallas Fed surveyed oil and gas executives about the use of AI, and over 56% said they had no plans to use it. A year later, when they were surveyed again, 74% said they were now using AI, and over two thirds planned to expand it. You have been building for these industries for a decade. What changed in the last few years that woke them up?

Gil: Before COVID, back in 2018, we had a major US refiner look at our technology and say, "This looks interesting. I'll adopt it, but I'm going to be the last company in the US to do it. I'm sticking to my old known methods." Over the last couple of years, that same US major has been moving pretty fast with us. We heard from multiple people that their board is mandating various teams to improve operations with AI. We are seeing this across the industry, all geographies. Boards and executive teams see other industries outside oil and gas moving a lot faster, and they are pushing their own teams.

Jay: Is there something specific the boards are pointing to? Labor, tech proof points, economic pressure?

Gil: It is the standard problems they are always trying to solve. Safety, reliability, margin improvement, compliance, and the competency of their workforce. Some of these assets can be 130 years old. Those are the major things, and they believe they can improve them with AI.

Jay: For 40 years these refiners have used linear programming to optimize operations, and you are using AI for non-linear optimization. Give me a specific process where linear programming makes a wrong prediction but AI catches it. What is the error delta in real dollars?

Gil: Take a major unit in oil refineries called the fluid catalytic cracker, or FCC. It takes catalyst, which is a nano-engineered metal, and puts it in reactors at high temperatures and high pressures. That lets you break cheap hydrocarbon molecules, the kind used for things like asphalt, into higher-value molecules that become jet fuel, diesel, and gasoline. Traditional controls help the operators run those reactors to a certain desired temperature. That target was set through a linear program. What AI like Imubit does is help the operator figure out what that temperature target should even be, as part of a larger operating strategy. That can be worth up to $10 million a year, and that is just one part of the plant. There are multiple other units. When customers saw this actually working, they started going to their leadership teams and boards. Some even spoke about it with shareholders.

Jay: What was the specific proof point that converted the skeptics? Something they were doing manually, or could not do manually, that changed their minds when they saw the AI do it?

Gil: They started seeing how the AI takes the market prices of jet fuel, gasoline, and diesel and figures out how to automatically adjust those temperatures and pressures so the unit yields the most profitable combination of molecules. Especially when your feedstock is not fully observable. You can have a ship coming in with crude from West Africa or South America, and even if two shipments come from the same source, they are slightly different in composition. Those are unobservable disturbances. The AI figures out how the feed is varying and makes very small changes that on average almost look like zero change, but they improve the value of that combined pool of molecules.

Jay: So it understands the difference between the inputs, then understands what outputs to create based on where the highest value in the market is.

Gil: Exactly.

Jay: This is the interesting part to me. You are not ripping out and replacing existing control systems. You are building a layer on top. Walk me through a deployment. What do you read from, what do you write to, and what always breaks when you integrate?

Gil: Let me tell you what is different about what we do versus other software and AI companies. We do closed-loop optimization. That means the AI does not just operate in the cloud. It is actually moving valves in an environment that can be dangerous if these things are not done safely. Most AI companies do not directly, physically touch these multi-billion-dollar production environments where safety and reliability are so critical. Picture it. You walk in and see a Honeywell or a Schneider Electric distributed control system, screens in green and black and white, and they could be 40 or 50 years old. Now imagine taking modern AI and integrating it into those things. That is the challenge in doing it safely.

Jay: What data do you actually need from those legacy systems? Real-time sensor feeds, pressure, temperature? What are the inputs?

Gil: You use the sensor data. Pressures, temperatures, flows, various lab analyzers. Those get collected in software called a historian. You use the data in that historian to train our models. Once the models are trained, we connect into the control system and keep collecting that data to keep training. And remember, closed-loop optimization means you are not just handing over a recommendation for someone to look at and maybe ignore. You have a neural network feeding right back into the control system, and in a safe, controlled way it actually changes the temperature every minute, 24/7.

Jay: A couple months ago Kevin Weil from OpenAI talked about a lab, might have been Berkeley or Stanford, where they connected OpenAI to lab equipment and could run simulations while moving sensors and settings based on the optimization of each recursive test. Instead of a human sitting there working it, you are taking that same principle and applying it at the scale of millions of barrels of oil. Is that the right way to think about it?

Gil: The key is that these plants are not ready for something like a large language model to start playing with them. You have to build your neural networks in a very special way to do that.

Jay: Some of this optimization is happening in a black box, and neural networks are largely black boxes. You have plant engineers who spent 30 years learning the physics, and now you are saying, trust the AI, it optimizes more precisely than you do. How do you earn that trust, and where is the friction?

Gil: It has been a painful question for us, and we learned a lot from that pain. We started by going after the hard problems first, like optimizing the FCC. Initially we did all the modeling ourselves, just to make sure the problem was solved and the value was generated. But it was a black box to the engineers at the site, and they were concerned about two things. One, being able to understand what the neural network had learned from the data. Two, being able to influence it, because sometimes the data is wrong. You can have a bad flow meter. Things change at the plant. They need to insert their first-principles understanding into it. Visibility into what the network has learned, and the ability to steer it toward learning the right things, has been the focus of most of our research over the past few years. Our neural networks are not open-source LLMs that are fine-tuned. That is not acceptable. We guarantee to our customers that the networks are only trained on their curated datasets. You start from a network that was never trained on anyone else's data, and the customer can literally train their AI model from scratch. Think about what that means. A chemical engineer, not even a data scientist, training their own AI model from scratch.

Jay: Is there a chicken-and-egg problem? To run a pilot you need access to their data. You cannot use publicly available stuff. It is their specific training data, but you have to convince them to give you access before you can show the value. How do you build the trust to get that first pilot?

Gil: Like anything, you start with the early believers. If you focus on one industry and do not spread yourself thin, you can get some of those proof points, and then others become open to testing. You obviously start with offline data. It has been, I would not say easy, but reasonable to prove these things to customers on their own data.

Jay: You mentioned the eagerness boards have to show value with AI. What is the timescale for these pilots? When do people expect to see value?

Gil: Traditionally it took us about six months from kicking off to getting a closed-loop application running. It still takes a few months, just because of the safety, methodology, and procedures involved. But we have shortened the time to value significantly, to weeks and even days, by starting to show them results on their own data in our platform. They can see what the models are learning and recommending even before it goes closed-loop.

Jay: You took an unusual hybrid talent approach, hiring oil and gas engineers, not just data scientists. Talk about the moment you realized top Valley talent alone was not going to win this market. What can only a domain expert see that a pure machine learning person cannot?

Gil: Think about the classic crossing-the-chasm strategy. Focus on a segment, go out and solve hard problems. It was clear early on that coming in with world-class data science alone was not enough to solve these multi-billion-dollar problems. My initial assumption was that I needed the best people from Honeywell, AspenTech, and AVEVA, the people who had served the market for a while. I would combine them with the data scientists and we would win. That failed miserably. The reason was that the folks from those industries found it hard, even harder than the customers, to rethink the solutions. To rethink the linear problem, to understand that AI can abstract away some of the things they had built their careers on. So I realized I needed something different. What I needed was up-and-coming process engineers from the target market. People from Exxon, Mobil, and Shell. People who were passionate about the chemical process itself and were curious about AI. I actually wrote a job description, and I had a line in it that said, "Come disrupt your own industry." I got a surprisingly high quality of early engineers who joined us. And I'll tell you about our first big customer testimonial. The customer said, "When Imubit came in, they brought 60 years of industry experience through that door. Not just data science." That gives you a sense of the impact.

Jay: You have landed seven of the ten biggest US refiners. Rank the core reasons they go with Imubit in order of importance. Technology, price, relationships, existing frustrations. Which ones did you think would matter more than they did, and which did you underrate?

Gil: These people have a lot of margin pressure, so the top reason is that we solved painful margin and efficiency problems they could not solve before. That is number one. Number two, we brought proven technology plus a white-glove domain-expert service, which was different from the incumbents. Number three, they were future-proofing their businesses. They were not exactly sure what they would get, but they moved their organization forward instead of backward. Now let me tell you about a wrong assumption I had, and it relates to the black box. I always assumed people don't like black boxes. But what I missed early on is that I thought refiners just wanted to run the plant, stay in their own domain, and would not be interested in building their own AI models. They just wanted the problems solved. That was a big mistake. We actually lost a key client because we did not have that capability yet. It took us years to build. As a result we made a lot more progress and realized it is actually key to unlocking scale, because we turned those engineers from black-box blockers into promoters through ownership. Now it is their models. I would not have expected engineers in refineries to want to build AI models themselves, but that is exactly the case.

Jay: That is a genuinely contrarian insight, because a lot of founders building for legacy industries assume these folks don't want to buy tech, they just want the job done. You are saying the more time you spend with these engineers, the more you see they deeply care about the tech. Why? Is it a business reason for owning the models, or is it about whose head is on the chopping block if something goes wrong?

Gil: It starts with a business case. They need to see the business case first, and the fitness of the technology to solve their problems. Otherwise they are not going to be interested in putting their hands on the models. But the more we scale and succeed at creating value with these models, the more they feel it is a critical part of their operation. One customer told me, "Look, we're a Fortune 100 company. At the time you had fewer than 100 people. Even if you had 1,000 people, even 10,000, I would not want to count on an external company to optimize my multi-billion-dollar assets and affect what my executives write in the annual shareholder report. We need that control. We want to leverage your technology, but we need to develop the skill set to do it ourselves."

Jay: What are your competitors getting wrong? To put names out there, Honeywell launched its own plant-wide optimizer, and AspenTech acquired a company in the space. What have you built that takes them three-plus years and, even with infinite resources, is hard to replicate? How long before they close the gap?

Gil: Those products are part of a layered pyramid in the process industries. At the top you have the linear program for the plant, then real-time optimization, then advanced process control, then distributed control systems. A lot of this has been perpetuated by these companies because they own all the layers. They want you to start at the lower layers and buy all the blocks and slowly build up. We assumed that was the case getting into the market too, but after a few years we learned it is a bit like a house of cards. You can build it in a perfect environment with no wind, here on my desk, but in reality it is like building a house of cards on a pit of sand. Reality is dirty, it varies quickly, and it breaks the hierarchy. Our technology is a lot more elastic. We do not need those rigid layers to be in place. Whatever they have, we adapt to it. And these incumbents suffer from the innovator's dilemma, because those siloed products are cash cows. They need to keep pumping cash, so we don't expect them to change course. We do see other AI companies trying to get into the market, so there is a window of opportunity, and we have to stay paranoid about competition. The real question is how you leverage generative AI into a solution like ours and reframe the job to be done accordingly.

Jay: How do you think about the resourcing gap? You have raised an impressive $100 million. Honeywell probably spends more than that annually just on food for its people. At what point did you feel you could beat these incumbents despite the gap, and what insight made you confident?

Gil: We built a cloud-native platform that lets non-data-scientists train deep learning neural networks from scratch, and we proved it across a lot of use cases. Those incumbents are still using desktop software. They are copying us by letting their users train neural networks, but those are closer to the networks of the '90s and 2000s. They are not modern deep learning networks leveraging cloud and GPU. We have major refiners that own all the latest and greatest from companies like Aspen, and they still use our models. This laser focus on one thing, closed-loop AI process optimization, is what we have done, and it is very hard. Once you prove beyond any doubt that you can do it, you can start to build on that, which is what we are doing.

Jay: You have talked a lot about focus, but focus has tradeoffs. From my research, the core refining market for you may be worth around $300 million, and your VCs expect a multi-billion-dollar outcome. How important is vertical expansion? If five years from now you are doing $500 million in revenue, but 85% is still from refining, is that a win or a loss?

Gil: Doing 85% in refining in five years would be a failure. We estimate the closed-loop process optimization market in refining is a bit higher, closer to $1 billion, but it is slow-moving, cyclic, and under pressure. It is not a collectively growing market. So vertical expansion is critical. We are doing this in cement, in chemicals, and we are starting in mining. We start the expansion directly, to learn from our customers, but to move faster we do it through partnerships. A few weeks ago we signed a strategic partnership with a company called Fuller [verify spelling], the market leader in automation for the cement industry, from machinery to hardware to software. Under different names they have operated for 130 or 140 years and they have a global footprint. That lets us move into new verticals without spending the years it took us to build proven domain expertise in oil and gas.

Jay: What makes a vertical ready for expansion? Similar equipment physics, similar margin structure, similar buyer behavior? Rank the hierarchy and give the rationale.

Gil: To simplify, you need a continuous process where decisions are being made by operators working shifts, 24/7. You need incentives to improve the process. We see similar buying dynamics across a lot of these plants. And you need a certain size. Cement is much smaller than refining, both in volume and in annual value of product produced, which is why working with partners like Fuller makes a lot of sense for us. But it always comes back to the top consideration, which is the North Star, and sometimes it is hard to see. Can you solve a problem the plant cannot practically solve with the alternatives? Some problems at some plants are just too simple for this to make sense.

Jay: On cement specifically, there has been a lot of verticalization. The largest producers have bought up smaller players, especially green cement. When you have that much consolidation, does it make it easier or harder for you to come in, run a pilot, train on their data, and build the models, when you are dealing with fragmented businesses inside one large parent?

Gil: We see all of it in oil too. Some companies are downstream or refining only, some are integrated across upstream, midstream, and downstream. So we are used to operating within very complex enterprises.

Jay: What is the metric you want to see 12 months from now to know cement was the right vertical?

Gil: We would like to see a large number of plants with initially smaller applications. Traditionally, in refining, everything is heavier and larger, operationally and technology-wise, and we start with large installations. In cement you just can't do that. This partnership accelerates the approach of having a higher number of smaller applications. So it is that number of applications, and we hope to be in dozens of plants within a few years through these partnerships.

Jay: I'd love to look forward. Five years from now, what will the industrial AI market have overestimated about what's possible, and what will it have underestimated?

Gil: That is always an interesting question. You have functions at the plant that have been perpetuated for literally decades, and people built entire careers deeply understanding those engineering disciplines and products. The assumption most people have is that those products keep evolving until AI agents start to operate them and gradually replace the people, and then the agents work together to run the plants the way humans do today. We don't agree with that. We think nobody is going to automate the work these people do in the next 10 years, maybe even 15. Instead, what emerges is a coordinated platform that lets these people, as people, not AI agents, coordinate their decisions across the plant. Like an operating system for the plant. Like a Salesforce for the plant. It brings all those cross-functional operating strategies into one system of record. That lets you benchmark how you are making cross-functional decisions, executing strategies, and managing constraints and objectives at the plant. Then you start deploying models to help those engineers and operators execute better, and to automate parts of it. What you see is a layer that first connects the people and gets them working together in a much more coordinated way. From there it becomes more natural and structured for AI agents to start automating more of the plant.

Jay: So the right way to think about it is like a command center that gives you visibility on what is happening, and then, based on the problem, decides whether a human should solve it or an AI agent should. But without the system-of-record layer, the command center isn't possible. You need the system of record first, then the command center, then the ability to deploy the right resource, human or AI. Is that right?

Gil: Yeah, that is a good way to think about it. Imagine everything people want to automate in sales without a Salesforce. That doesn't make sense, and that is what is happening in these plants right now.

Jay: Gil, this has been a pleasure. Continuous processes, millions of units of volume, margin-constrained environments where the cost of error is so much higher. It makes it all the more impressive that you and the team at Imubit have built this trust with your customers. I'm excited to have you back to see how many of these predictions come true. Thank you for joining me on Climb.

Gil: Thank you so much for having me, Jay. I look forward to it.

Pull quotes

"Our AI doesn't just operate in the cloud. It is actually moving valves in an environment that can be dangerous if these things are not done safely."
"You start from a neural network that was never trained on anyone else's data, and the customer can literally train their AI model from scratch. A chemical engineer, not even a data scientist."
"I needed the best people from Honeywell, AspenTech, and AVEVA. I would combine them with the data scientists and we would win. That failed miserably."
"Doing 85% in refining in five years would be a failure."
"Nobody is going to automate the work these people do in the next 10 years, maybe even 15. Imagine everything people want to automate in sales without a Salesforce. That doesn't make sense, and that is what is happening in these plants right now."