# Operator competence when AI runs the plant Operator competence with AI in the control room means the ability to run the unit by hand when the model is unavailable or wrong, to spot when it is wrong, and to stop it. Deskilling is the slow loss of that ability once the AI does the routine work. The AI control room operator still needs those skills, assessed and given regular practice, whether the operator is in the loop, approving each move, or on the loop, supervising a model that acts. Automation bias, the tendency to accept the model's output without checking, grows as the skills fade. This page is part of [AI in process safety management](https://www.autonomous-operations.ai/ai-process-safety-management/). ## Is deskilling a recognised risk of AI in operations? Yes. The HSE Science Division collected 247 AI use cases across 14 sectors and listed among the risks "over-reliance on AI safety systems, workforce deskilling" ([Chambers et al., Hazards 35, 2025](https://www.icheme.org/media/29576/hazards-35-paper-11-chambers.pdf)). In its [statement of 12 June 2026](https://press.hse.gov.uk/2026/06/12/hses-regulatory-approach-to-artificial-intelligence-ai/), HSE said health and safety law already requires a risk assessment for uses of AI that affect workplace safety, and that it is researching the effects of cognitive offloading on human performance in major hazard industries. Vendors present the same shift as a benefit. Honeywell says its Experion Cognition agents act "on behalf of the operator" on anomalies and help "less experienced operators to run plants with comparable knowledge as that of a seasoned veteran" ([Honeywell, 9 June 2026](https://www.honeywell.com/us/en/press/2026/06/honeywell-introduces-experion-cognition-to-deliver-autonomous-control-room-operations-for-borouge-international)). The person who must take over when the agent is wrong may then be the one with least experience of running the unit without it. ## What must an operator still be able to do with the AI down or wrong? Four things, each of which should appear in the role profile. - Know the model's validity range and recognise when the process has left it. - Decide to stop the model, with clear authority to do so. - Run the unit in the fallback mode, including the manual tasks the model took over. - Handle an upset with the model disabled or behaving oddly. At ENEOS Materials Yokkaichi, an FKDPP reinforcement learning model took over two valves that operators had controlled manually with a heavy workload. Yokogawa's account quotes the production team: "I received several calls on my cell phone from operators asking if it was OK to stop the AI", and "if things go wrong, we can simply stop the AI and go back to normal DCS operation" ([Yokogawa, ENEOS Materials](https://www.yokogawa.com/us/library/resources/references/successstory-eneos-materials/)). The second quote holds only while operators can still run that normal DCS operation. The fallback and the switch-over belong in procedures, as set out on [AI in process control and the DCS](https://www.autonomous-operations.ai/ai-process-control-dcs/). ## How do you assess competence that is rarely used? Assessment has to test the manual role, as well as the use of the AI's screens. Practical assessment on a simulator or in a planned manual period, with scenarios where the model gives plausible wrong advice, shows whether the operator detects it. The less often a skill is used on the job, the shorter the reassessment interval should be. ## How do manual skills fade, and how do you keep them? They fade through lack of use. Keeping them takes planned practice: periodic shifts or periods with the model switched off, simulator sessions on start-up, shutdown and upset scenarios, and drills on taking the model out of service. Recording them in the competence system shows when practice lapses. ## What changes in staffing and organisational MOC? OSHA treats changes in "staffing levels, staff experience, or contracting out" that affect a covered process as subject to management of change ([OSHA, 31 March 2009](https://www.osha.gov/laws-regs/standardinterpretations/2009-03-31-0)). A model that takes over monitoring or control can be such a change when it leads to fewer operators or a less experienced shift. The review should ask which tasks move to the model, who keeps the competence to do them by hand, and whether the shift can manage an upset with the model down. The procedure side is covered on [AI management of change](https://www.autonomous-operations.ai/ai-management-of-change/), and permit judgements that move to AI on [AI in control of work](https://www.autonomous-operations.ai/ai-control-of-work/). In the competence management system, that means adding the AI fallback tasks to role profiles, the model's limits to training, and the manual practice to the assessment plan. I was operations manager and then deputy general manager of the In Amenas gas processing joint venture in Algeria, with more than 1,000 people, and later built a national operations team, including 48 apprentice technicians, for an FLNG start-up. Before that I led operations at an upper-tier Seveso plant. [M-OPS OMI](https://www.autonomous-operations.ai/omi/) checks whether your competence system and organisational MOC cover the model. The [autonomous operations briefing](https://www.autonomous-operations.ai/#how-do-i-work-the-m-ops-service-range) applies it to one of your units. ## What you can expect from XO Autonomous Operations - A check of whether your competence system and organisational MOC cover the AI in use. - The AI fallback tasks added to role profiles, and the model's limits added to training. - An assessment and practice plan for the manual role. ## Questions ### How do we stop operators losing skills when AI runs the plant? Write the manual skills into the role profile, assess them, and give operators regular practice running the unit with the model off, in the simulator or in planned periods on the plant. Include recognising when the model is outside its valid range and the authority to stop it. Review staffing through organisational MOC before tasks move to the AI.