Where does AI already write to the process?
- Aramco Fadhili gas plant, Saudi Arabia. On 29 October 2025 Yokogawa reported that several AI agents built on its FKDPP reinforcement learning algorithm “directly and autonomously control” the acid gas removal unit through the CENTUM VP control system, a project still under detailed evaluation (Yokogawa).
- ENEOS Kawasaki refinery, Japan. An AI system jointly developed by ENEOS and Preferred Networks has run an atmospheric crude distillation unit autonomously since January 2024, adjusting 13 valves at the same time (Preferred Networks, 24 May 2024).
- ENEOS Materials, Yokkaichi, Japan. FKDPP was formally adopted on 30 March 2023 for direct control of a distillation column (Yokogawa) producing butadiene (Yokogawa success story).
- Big West Oil, North Salt Lake City, United States. A refinery and Imubit case study describes work that started as closed-loop process optimisation, first on the FCC unit (PTQ Q3 2025).
What changes when AI agents run the control room?
AI agents now handle part of the control room’s work. At Fadhili, several FKDPP agents control one unit. At Borouge, Honeywell’s Experion Cognition agents act “on behalf of the operator” on anomalies (Honeywell, 9 June 2026). Process safety management then has to answer four questions.
- Scope. Which tags, units and actions each agent may touch, written down and changed only under management of change.
- Authority. Who can stop an agent, and how the operator takes back control in one step.
- Traceability. Every agent action is logged with its inputs, so it can be reviewed, reversed and investigated.
- Competence. The tasks the agents took over are the ones operators stop practising. See operator competence when AI runs the plant.
Moving tasks from operators to agents is also an organisational change. The wider picture, level by level, is on autonomous operations.
What forms does AI take in and around the DCS?
- Soft sensors that infer a composition between analyser readings.
- Models inside APC or real-time optimisation, setting targets for the regulatory layer.
- Operator advisory and AI copilots.
- Alarm ranking or suppression.
- AI features arriving with a DCS or historian upgrade.
- Closed-loop control, where the model moves set points without operator confirmation.
Is the operator in the loop or on the loop?
In open loop the model advises and the operator decides. In closed loop the model acts and the operator supervises, which is human on the loop. Hazard analysis, procedures and competence then need review, and the move is a change in its own right. In the Linmara case it was made without management of change, and a soft sensor set amine circulation from a feed it had never seen.
What has to be managed?
Ownership. Each model needs a named owner on site who approves retraining and can take it out of service. A vendor is not an owner.
Validity range. The model should only act within the feeds and instrument states it was validated for. Outside it the DCS should tell the operator and fall back to a defined mode, such as the conventional controller. Where it writes set points, its constraints must match the operating limits.
Running without it. Operators must still run the unit when the model is unavailable or wrong, which needs procedures and practice.
Alarms. Alarm ranking changes what the operator sees first. It falls under the site’s alarm philosophy and the lifecycle of IEC 62682.
Updates. A retrain or a vendor release that changes AI behaviour goes through management of change. In the Quillmere case a monthly vendor update reranked a reactor temperature alarm outside management of change.
Cybersecurity. The ISA/IEC 62443 series rests on responsibility shared between asset owners and suppliers. Each new data path into or out of the DCS needs assessing before it opens.
Should AI be part of the safety instrumented system?
IEC 61508 and IEC 61511 contain no AI-specific guidance, and ISO/IEC TR 5469 gives guidance only. Keep the SIS independent of any AI in the basic process control system, and check that an AI layer does not erode the independence your LOPA assumes. See AI and safety instrumented systems.
Does the EU AI Act make a DCS model high-risk?
Annex III point 2 classes as high-risk AI systems intended as safety components in the management and operation of critical digital infrastructure, road traffic, or the supply of water, gas, heating or electricity. A gas or power operator using AI in control should record whether each model is a safety component, and why. The separate Machinery Regulation (EU) 2023/1230 applies from 20 January 2027 (corrigendum, OJ L 169, 4 July 2023); the dates are set out on EU AI Act and Machinery Regulation dates.
The wider picture is on AI in process safety management. To see where AI sits in your control system, start with an autonomous operations briefing.