Across all sectors the picture is the same. In PwC’s 29th Global CEO Survey of 4,454 chief executives in 95 countries and territories, more than half (56%) said their company had seen neither higher revenues nor lower costs from AI; only one in eight (12%) reported both (PwC, 29th Global CEO Survey, January 2026).
In a high-hazard operation an AI system cannot reach production until management of change, risk assessment and barrier management can accept it; where they have no route for it, it stays a pilot.
Why is AI harder in high-hazard operations?
The US National Institute of Standards and Technology calls AI “a uniquely challenging technology to deploy and utilize” for any organisation (NIST AI 100-1, 2023). Its reasons: the data a model learns from can change over time, sometimes significantly and unexpectedly; AI systems and the settings they run in are complex, so failures are hard to detect and respond to; and AI is socio-technical, shaped by who uses it and how.
In an operation that manages major accident risk, each of these weighs more.
- Data. A process changes with feed quality, equipment wear, catalyst ageing and every modification. Upsets and rare events, the conditions that matter most, are the ones a model has seen least.
- Failure. A wrong output can look plausible. Where the process safety time is minutes, an error that is hard to detect may not be detected in time.
- People. Operators act on what the screen tells them. Reliance grows, manual skills fade, and the control room can stop questioning the system on the day it matters.
- Assurance. Process safety rests on systems that are specified, tested and proof-tested. NIST notes that testing standards for AI are underdeveloped, and the functional safety standards the industry applies, IEC 61508 and IEC 61511, contain no AI-specific guidance.
The United States’ AI Action Plan of July 2025 sets out to “Remove Red Tape and Onerous Regulation”, and still expects AI in safety-critical applications to be “instrumented to detect performance shifts” (The White House).
Elsewhere, an AI error usually means a poor decision or a cost. In a high-hazard operation it can weaken a barrier against a major accident. Many operators therefore keep AI at pilot stage, or deploy it without the controls their safety case assumes.
How does an AI model move from pilot to scale?
Through stage gates, each passed with its own management of change review, because each step gives the model more reach into the process. As an illustration, take a model that recommends setpoints for a distillation column.
- Pilot. The model runs in parallel and nobody acts on its output. MOC records that it is outside the decision path and approves its data feeds and any connection to the control network. Once operators start acting on it, it is at the next gate.
- Advisory. Operators see its recommendations and decide. This is first deployment into a decision path. MOC defines the operating envelope, what the model cannot see, when to disregard it, and the training operators complete before start-up.
- Closed loop. The model writes setpoints to the process control system. MOC reviews write limits, interaction with alarms and safety instrumented functions, whether credited protection layers stay independent, and how operators take back control. The hazard study is revisited, and organisational MOC covers the monitoring tasks that move to the model.
- Fleet. Each further unit or site is a new change, with its own equipment, instruments and operating envelope. Retraining for that unit, and each vendor update afterwards, returns to MOC.
What does this mean for an operator?
AI delivers when it starts from the operator’s own aims and is built into the management system and processes that already run the plant: management of change, risk assessment, competence, procedures, barrier management, monitoring and records. That is how the M-OPS service range works.
If AI is not delivering in your operation, or you are not sure where to start, contact me.