PTQ Q4 2026 Issue

The real work begins after the alarm

AI and IoT have enhanced the ability to generate better signals, but nothing changes without an owner who can close the loop on exceptions

Dinesh Varadharajan Kissflow

O il and gas companies have invested heavily in their data capabilities over the last few decades. Data has become more reliable, and dashboards give a clearer picture than ever before. However, as the investments in better data stack up, organisations still have not crossed the line to show real ROI. Plants and refineries know when something is wrong, but this knowledge has not translated into true operational savings. The answer is not in better models, improved sensor cov- erage, or more integrations. Changes in compressor vibra- tion can already be detected within seconds. Distributed control systems (DCS) give highly refined alarms. AI can be used to compare patterns against years of data and rank the likely causes. However, to get the true value of these insights, people are needed. More specifically, owners are needed. Owners who can validate signals against years of experience. Owners who can navigate relationships with operations, process engineering, maintenance, inspection, and health, safety, and environment (HSE) to reach a solu- tion. Owners who understand fieldwork, permits, contractor availability, and production conditions. The perfect blend of technology and people is in exception management. It is the way to close the final gap between signal and response. How organisations handle exception management separates those who are simply throwing money at problems from those who can scalably solve them. Exception management in a refinery Exception management is the operating discipline that con- verts a deviation from an expected state into an owned, time-bound, risk-based action and keeps that action open until the result is verified. It is different from alarm manage - ment. The International Society of Automation (ISA) defines an alarm as an audible or visible indication of an equipment malfunction, process deviation, or abnormal condition that requires a timely operator response. A poor alarm system burdens the operator and weakens situational awareness. However, an exception can continue long after the alarm has cleared or acknowledged. Some exceptions do not start with an alarm at all. An unexpected turnaround finding is an exception. A few examples include a temporary repair approaching expiry, a repeated seal failure, or a laboratory result outside specification. The alarm tells the operator that attention is required. Exception management carries the matter across functions until the refinery has a defensible outcome.

The outcome may be a repair, a controlled rate reduction, a new monitoring requirement, or a decision that no interven- tion is required. The important point is that the decision has an owner, a technical basis, a time limit where appropriate, and evidence of closure. This is where better data breaks down. A dashboard can show that a condition exists, but it cannot make a named person accountable for resolving it. Solving the response problem AI is useful in refinery operations. It can find weak patterns in large data sets, identify equipment behaviour that fixed thresholds miss, and summarise operating history. However, a model does not carry operating authority. Suppose a pre- dictive model identifies an increased probability of fouling in a crude preheat exchanger. The model may be correct. The refinery still has to decide what to do with that information. Can the unit continue until the planned cleaning window? Does the current crude slate change the assessment? Is the differential pressure trend consistent with the model? What does the laboratory data show? Will an operating adjust- ment affect furnace duty, hydraulic limits, product quality, or downstream constraints? The model sharpens the signal, but it cannot answer these questions. AI drastically improves detection and decision support in ways humans cannot. However, the governance layer turns the recommendation into an owned, recorded action within the refinery operating system. The National Institute of Standards and Technology (NIST) makes the same principle clear in its AI Risk Management Framework (AI RMF 1.0).¹ Trustworthy use depends on governance, defined responsibilities, documentation, measurement, and management across the life of the system. In a refinery, these are not abstract AI controls. They become shift roles, approval limits, operating procedures, management of change (MOC) requirements, permit condi- tions, and verification steps. The strongest AI architecture in a refinery is not the one with the most models. It is the one that knows where a model stops, and where accountable operations begin. Systems should carry the work Refineries already have strong specialist systems such as DCS, laboratory information management systems, comput- erised maintenance management system (CMMS), and per- mit-to-work systems. The problem is not that these systems are not useful. The problem is that an operational exception

PTQ Q4 2026