Gilad Cohen tells a useful failure story. Early in Imubit's life, he assumed refinery engineers wanted one thing from artificial intelligence: a difficult problem solved. Build the model, improve the economics, show the result. In a sector where a single unit optimization can be worth millions of dollars a year, that sounded like a sturdy value proposition. It was also incomplete.
The engineers wanted to understand the models. They wanted to train neural networks on their own plant data and inspect how the system reached its decisions. Cohen has said the first approach “failed miserably.” A key client was lost. Imubit then spent years making the platform more usable by the domain experts who would remain accountable when software crossed the line from observing a process to influencing it.
That correction is the center of Cohen's story. It makes him a particularly instructive figure in the current AI cycle, when software companies can confuse technical capability with permission. In a refinery, permission is concrete. The system reads temperatures, pressures, flows and lab results. It may write setpoints back to control infrastructure that has been running for decades. A bad recommendation is inconvenient. A bad control decision can be expensive.
“I would not expect that engineers and refineries would like to build AI models by themselves, but that's exactly the case.”Gilad Cohen, speaking about an early product lesson
Make the complicated tractable, without pretending it is simple
Cohen's route to the control room began with mathematics. At Ben-Gurion University of the Negev, he completed undergraduate degrees in mathematics and electrical engineering, followed by a master's degree in electrical and computer engineering. His graduate work dealt with nonlinear, non-convex, multidimensional pattern-detection problems. The basic move was elegant: project the problem into a space where classical linear detection tools could address it without simply flattening away the original complexity.
In 2008, he co-authored papers on detecting, segmenting and registering elastically deformable objects. He also received an Israeli Ministry of Defense award that year for outstanding research and development of an aerospace system. The subject matter was removed from refinery economics, yet the mental habit is recognizable. Find a representation that makes a hard system operable. Keep the constraints. Use established tools where they work.
The operating loop Cohen is trying to close
After university, Cohen moved from algorithms into company building. In 2010 he founded Cigol Digital Systems, which developed edge cybersecurity software for data-center systems. He led the business until its acquisition by Mellanox Technologies in 2017. Cybersecurity at the edge and industrial control share a productive tension: useful software must participate in a live system while respecting boundaries it cannot casually cross.
By the time that acquisition closed, Cohen had already co-founded Imubit with Nadav Cohen in 2016. The new company took on an industrial problem that sits between software, process engineering and organizational behavior. Refineries and chemical plants contain many linked units. Operators are balancing yield, energy use, equipment constraints, feed composition, product specifications and market prices. Optimizing one piece can move the problem elsewhere.
A prediction is only valuable when the operation changes
Much of enterprise AI stops at a dashboard. A model predicts an outcome, highlights an anomaly or produces a recommendation. The human user must still translate that signal into a decision, then translate the decision into a control change. Cohen's argument is that this last mile contains much of the unrealized value.
Imubit's software is designed to sit above existing distributed control and advanced process-control systems, rather than requiring plants to replace them. It learns from historical and live plant data, models nonlinear relationships, and can adjust operating setpoints within defined constraints. The company calls the approach Closed Loop AI Optimization. The phrase matters less than the discipline underneath it: read what the plant is doing, connect the behavior to economics, act within boundaries, observe the result, repeat.
In 2021, Imubit announced $50 million in total funding, including a $30 million growth round led by Zeev Ventures and Insight Partners. In that announcement Cohen pointed to a customer-reported annual margin improvement of $10 million from optimizing a complex refining system. More recent company material says the platform has passed 100 deployments and works with seven companies included in a ten-member ranking of US refiners.
Those numbers explain why the market pays attention. They do not explain why a model stays online. For that, Cohen returns to people. Refinery engineers carry deep knowledge of equipment behavior and local operating history. Operators work inside a rhythm of alarms, handoffs and constraints. If AI arrives as an inscrutable verdict, it asks those people to surrender agency while leaving responsibility exactly where it was.
Cohen's response has been to make model building and monitoring accessible to non-data-scientists, ground the system in engineering constraints, and emphasize transparency. This is not transparency as a compliance ornament. It is a product mechanism. The engineer can challenge a relationship that looks wrong. The operator can see how a proposed move relates to the plant's state. The model becomes an object around which operations, engineering and commercial teams can coordinate.
“Talk of AI is everywhere, but in the energy industry, a lot of what’s being called AI is simply a re-spin of traditional linear process optimization.”Gilad Cohen at an Imubit technical program
Automation works when expertise gains leverage
Cohen is careful to distinguish the specialized models used in plant operations from the probabilistic language models dominating public attention. A refinery model is trained on structured numerical telemetry from the site: temperatures, pressures, flow rates, laboratory measurements, thermodynamic limits and physical constraints. Its job is not to produce plausible text. Its job is to represent a process closely enough to support a bounded operating decision.
This distinction has become more important as generative AI enters industrial software. In 2026, at an Imubit event in Houston, Cohen argued for open and interoperable industrial systems rather than a single vendor's all-encompassing architecture. The position fits the rest of his career. A working plant already has control systems, historians, planning tools, specialized expertise and a long memory. New intelligence has to join that ecosystem.
A practical trust checklist
Can experts inspect the model? Can the system operate inside explicit engineering limits? Can a human understand, challenge and revise the operating strategy? Cohen's public arguments repeatedly return to these three questions.
The approach is also expanding beyond unit optimization. In May 2026, Imubit and catalyst company Ketjen announced pilots that combine live refinery data with catalyst expertise. Cohen described a platform that continuously learns how complex units operate, giving teams earlier visibility into performance changes and clearer guidance. The partnership is a small example of his larger architectural idea: domain knowledge should connect to live operations without being trapped in a separate report.
Cohen's calendar reflects how much of this work happens in rooms, not release notes. After returning from the Global Refining and Petrochemicals Congress in India in 2024, he wrote about meeting leaders from major oil companies and about Imubit's local partnership with Tridiagonal Solutions. It was his second visit to India in three months. What caught his attention was not a futuristic plant tour, but the “growth mindset, rapid execution, and openness” of refiners, cement producers and steelmakers. A year later, at gatherings in Athens and Munich, he returned to the recurring questions: Who controls the data? Can technical teams trust the model? Does the investment connect to the economics of the plant? The itinerary changes. The adoption problem remains stubbornly local.
There is a workforce argument here too. Process industries face the retirement of experienced engineers and competition for younger technical talent. Cohen presents AI as a means of carrying knowledge forward, improving training and giving new engineers more powerful tools. The ambition is not to erase the operator from the picture. It is to let a new generation explore process relationships faster while making the experience of veteran staff legible inside the system.
That is why the early customer loss matters more than a tidy founder anecdote. It revealed the real product. Imubit was not merely selling a neural network. It was negotiating a new division of labor among software, engineers, operators and existing controls. The technical model had to learn the plant. The company had to learn the people.
Do not confuse an answer with adoption
Cohen's work offers a portable lesson for anyone building AI for expert users. Start with consequence. Who is accountable when the system acts? What context do they possess that the model does not? What would let them test the system without surrendering authorship of the decision?
Then respect the installed world. Plants do not become blank slates because a new model arrives. Neither do law firms, hospitals, banks or logistics networks. The older systems may be awkward, but they encode investments, routines and institutional knowledge. Adoption is often an integration problem before it is an intelligence problem.
Finally, treat a failed assumption as information. Cohen's first idea about refinery engineers was plausible. The customer showed him where it broke. His useful move was not defending the premise, but rebuilding around what the users needed to remain capable and responsible.
Industrial AI will continue to become more autonomous. Cohen's bet is that autonomy earns its place through controllability: specific models, visible reasoning, explicit boundaries and experts who can shape the system. The software may touch the controls. The people still set the purpose.