August 19, 2026 Five Minutes Instead of Forty: How Industrial AI is Rewriting the Economics of Refinery Safety

AI helps refinery teams respond to complex plant-floor risks that regulations alone cannot anticipate, supporting faster, safer decisions.

Regulations describe the correct response to known situations. But real problems at oil refineries begin with a combination of deviations that no regulation anticipates. Yeldos Nugumarov, Executive Director and Production Director of Aktau Petroleum LTD and co-founder of PetroLogic, explains exactly where the decision-making chain breaks down and why AI on a plant floor works not instead of people, but alongside them.

The Problem That Regulations Cannot Solve

Oil refining remains one of the most complex and heavily regulated industries in the world. Dozens of control systems, multi-layered protocols, mandatory personnel certifications. And yet, according to industry data, 40 to 60 percent of incidents at refineries are still linked to human factors. A paradox? Only at first glance. Regulations cover known scenarios, while most real failures originate where several minor deviations converge, none of which appears critical on its own.

The primary damage to the industry comes not from headline-grabbing disasters but from routine losses that accumulate day after day. A 2 to 5 percent drop in product quality, fuel overconsumption, forced load reductions, an unplanned shutdown lasting 4 to 8 hours. Individually, each episode seems tolerable. But multiplied by frequency, they add up to hundreds of thousands of dollars in annual losses at a single facility. The problem is particularly acute for small-scale refineries, where resources for multi-tier safety systems are significantly more limited, yet the cost of downtime relative to revenue is equally painful.

In recent years, the industry has increasingly looked toward artificial intelligence as a tool capable of closing precisely the gap that neither regulations, nor training, nor discipline can cover: the gap between the appearance of an early signal and its correct interpretation. But between the idea and a functioning system on a real plant floor lies a distance that only a few have managed to cross. One of them, Yeldos Nugumarov, covered that distance over four years, from a pilot project at his own facility to deployments across several industrial sites in Kazakhstan.

Yeldos Nugumarov is not someone who came to industrial AI from IT. He came from the plant floor. As Executive Director of Aktau Petroleum LTD, he built and launched an oil refinery from the ground up, oversaw its modernization and capacity expansion, established the full operational cycle, and brought the facility to stable production. Throughout the entire period of operation, the refinery maintained a consistently strong track record of customer satisfaction. Product quality was separately recognized through letters of commendation from Jan De Nul Kazakhstan LLP, a global leader in dredging operations, and PTC-Holding, an international provider of integrated logistics services in the oil and gas and container sectors. For a small-scale refinery, that level of client trust speaks louder than any award.

But oil refining was neither the first nor the only industry in his career. Over fifteen years, Yeldos Nugumarov progressed from gold mining to executive roles in marketing and business development at the country’s largest shopping and entertainment complex. His career began at Syntas Gold Mining LLP, continued in fuel retail at Helios, then moved to Home Mart, part of the Eurasia RED holding company with Aport Mall in its portfolio, where he led marketing and business development at the executive level. It was precisely this cross-industry experience, from industrial engineering to retail data analytics, that later formed the foundation of his approach to industrial AI.

In 2022, already understanding the pain points of small-scale refineries from the inside, Yeldos Nugumarov co-founded PetroLogic with Artem Kolmogorov and began building what did not yet exist on the market: proprietary AI systems for industrial facilities. He is a member of the Society of Petroleum Engineers (SPE), co-author of peer-reviewed publications in Energies and Eng (MDPI, 2026) and in Neft i Gaz (2026), Kazakhstan’s leading petroleum industry journal, and author of an expert article on industrial intelligence in Petroleum Journal (2025). The research on paraffin crystallization in high-wax crude oils from the Uzen field was carried out under a grant from the Science Committee of Kazakhstan’s Ministry of Science and Higher Education.

Yeldos, human factors are cited as the cause of the majority of incidents at refineries. But what exactly lies behind that statistic? What does a typical error look like at a small-scale plant?

A typical error at any refinery is not a single wrong step. It is a chain of five or six links, each of which looks tolerable on its own. It all starts with an early signal: a minor deviation in temperature, pressure, flow rate, or vibration. The operator either fails to notice it amid the stream of data or interprets it as normal process noise, and that is a rational decision under conditions of incomplete information. Then comes a delayed response: on average, 5 to 40 minutes pass between the appearance of a deviation and the first action, sometimes longer. After that, a locally logical but systemically flawed decision is made: one parameter is adjusted without tracking that the problem has already migrated to an adjacent unit or created a load on downstream processes. The result is not a catastrophe, but a loss: a 2 to 5 percent decline in product quality, fuel overconsumption, reduced throughput, or an unplanned shutdown lasting 4 to 8 hours. Multiplied by the frequency of such episodes, that amounts to hundreds of thousands of dollars per year at a single facility. And that is before accounting for the management domino effect, where one deviation cascades into shipping disruptions, accelerated equipment wear, and an overloaded next shift.

Can you describe a specific case where your AI system prevented a potential incident? We are interested in the full cycle: from signal to outcome.

I will give you one illustrative case. The system detected not an isolated threshold breach but a dangerous combination: a deviation in the temperature profile of a distillation column, an accelerating rate of pressure increase, and indirect signs of unit destabilization. Three parameters, each of which a duty process engineer could often explain individually without connecting them into a single picture. The system produced not just an alarm but a structured assessment: the specific unit, the nature of the deviation, the probable cause, and the expected trajectory over the next 20 to 30 minutes if the current operating mode was maintained. The crew verified the section, adjusted the operating parameters, and conducted an unscheduled equipment inspection. By my estimate, without that warning, the probability of a forced shutdown within 1.5 to 2 hours was 60 to 70 percent. The cost of such a shutdown at that particular section runs to approximately $80,000 to $100,000 per day, factoring in lost output and recovery operations. The key point is this: the system did not make the decision. It compressed the time from signal appearance to correct situational understanding from 30 to 40 minutes down to 5. That window is often what determines whether the outcome is a routine adjustment or an expensive shutdown.

You use the term “red flags.” What does that look like in practice for an operator, a process engineer, or a shift supervisor? And how do you address the problem of alert fatigue?

“Red flags” are not simply a notification system. They are a differentiated delivery of information tailored to a specific role and a specific decision-making context. An operator needs a short, actionable signal: what exactly is deviating, how urgent it is, what to check first. No more than three lines. A process engineer needs context: the trend over the past 2 to 4 hours, interrelated parameters, the probable cause, the impact on operating conditions. A shift supervisor needs the production-level picture: the risk to output, quality, and section stability through the end of the shift. The same alarm delivered identically to everyone is not a risk management system. It is a noise generator. The second principle is strict control over signal volume. If the system generates more than 15 to 20 notifications per shift, operators begin ignoring them. In medicine, this phenomenon is well documented and known as alarm fatigue. We address it through three mechanisms: prioritization by criticality level, consolidation of related deviations into a single incident, and mandatory explainability for every signal. When an operator understands why this particular signal demands attention right now, he responds. When he does not, he ignores it.

In your experience, where does the correct boundary between AI and a human being lie on a plant floor? Which decisions can be handed to a machine, and which cannot?

The boundary runs along the cost of error and the irreversibility of consequences. Where a decision is analytical and reversible, AI can operate with a high degree of autonomy: risk ranking, trend diagnostics, searching for analogues in historical data, inspection prioritization, operating regime recommendations. This is routine work that humans perform worse than machines in both speed and consistency. Where a decision affects the operating regime and requires engineering judgment, AI recommends and the human decides. A hypothesis about the probable cause of instability, a choice between correction scenarios: these require specialist verification. Where the cost of error is measured in human safety and equipment integrity, there is no AI autonomy whatsoever. Emergency switching, shutdowns, restart after intervention, overriding safety interlocks: this is a zone where the final decision belongs to a human alone, with no exceptions. Not because AI is not smart enough, but because accountability for these decisions cannot be delegated to an algorithm. AI is a second layer of thinking. Fast, tireless, immune to the attention drift that sets in at four in the morning. But it is not the plant dispatcher.

How did engineers and operators react to the introduction of the AI system? Was there resistance, and how did attitudes change over time?

The initial reaction splits into two camps, and this is predictable. Those working under chronic data overload see the system as a relief tool and adopt it quickly. Those who have built their own framework of judgment and interpretation over the years perceive a new tool as a challenge to their expertise, and that is a rational position, not mere conservatism. At the outset, resistance concentrates around three points: fear of intensified oversight from management, doubts about the quality of the system’s recommendations, and apprehension about being placed in a position where the algorithm “knows better.” A deployment built on the logic of “AI is smarter than you now” is doomed. The opposite logic works: the system amplifies your expertise, takes over routine signal processing, and frees you for decisions that require engineering judgment. The turning point comes after 2 to 3 specific cases where the system caught something the human missed and it had measurable consequences. After that, attitudes shift dramatically. An interesting observation: the most experienced specialists, once they are convinced of the system’s value, become its most consistent advocates. Precisely because they understand better than anyone what a 20-to-30-minute advantage in early deviation detection is worth.

What percentage of incidents at refineries is attributable to human factors, and what share of those can realistically be addressed by AI today?

Industry data for oil refining cite figures of 40 to 60 percent of incidents where human factors are a direct or contributing cause. But I would add a nuance: what matters more than the percentage is the mechanism. Human error rarely looks like a gross mistake. More often it is a systemic failure in the perception chain: late recognition of a signal, misinterpretation under conditions of incomplete information, delayed escalation, incomplete shift handover, or application of an outdated behavioral model to changed conditions. This is not incompetence. It is the predictable limitations of human attention under load. AI closes precisely this zone. It does not get tired, does not lose concentration at the end of a night shift, and is not prone to the tendency to confirm an already-formed interpretation when reading trends. 

Based on our experience, early deviation detection systems reduce the time to correct response by 40 to 60 percent and lower the frequency of unplanned shutdowns by 15 to 25 percent within the first year of operation, provided deployment and personnel training are handled properly. That is the typical first-year result. On individual sites, once the system has matured on real production data, the effect can go much further: at one regional mini-refinery, unplanned shutdowns were eliminated entirely in 2024, with an economic impact exceeding 20 percent for the year. But AI does not replace discipline, protocols, or a culture of safety. It functions as a multiplier where the foundation already exists.

In conversations about industrial AI, two poles typically dominate: enthusiastic promises of fully autonomous production and the skeptical refrain that “it will never work at our plant.” Yeldos Nugumarov occupies a third position, and what makes it compelling is that it grew not from theory but from the daily operation of a real refinery. The machine calculates faster, does not tire, and does not lose focus at four in the morning. The human bears responsibility, makes the final call, and grasps context that cannot be reduced to data. The task is not to choose one over the other but to draw the boundary between them with precision.

That boundary, in Yeldos Nugumarov’s framework, is not philosophical but engineering: three levels, each justified by the cost of a potential error. This approach does not promise a revolution, but it delivers something more valuable: it works. Plants confirm the results. International clients send letters of commendation. Peer-reviewed journals publish the research.

Yeldos Nugumarov says the next step is a project of global scale at the intersection of oil refining and artificial intelligence. Given that in four years he has progressed from a pilot on his own small-scale refinery to deployments across multiple industrial facilities and peer-reviewed publications, the industry has good reason to watch that next step with interest.

Subscribe to Industry Today