Havtil survey: operators lack oversight of their AI systems
23 October 2026.
Havtil, the Norwegian offshore safety regulator, surveyed operating companies on how they manage AI: oversight, risk management, follow-up of suppliers and assessment of safety effects. Practice varies widely. Some companies have registers, roles and processes to identify and monitor AI systems; others have little view of the AI being developed, bought or used, including AI built into suppliers’ products. Several companies said they had no safety-relevant AI while describing AI used in well planning, maintenance and analysis of operating data. Havtil notes that decision-support tools can be treated as safety-neutral even though they shape decisions and priorities, and that the regulations require indirect risks to be addressed too. It will use the findings in dialogue, guidance and supervision (Havtil).
What it means for operators: an up-to-date register of AI in use, including AI inside suppliers’ products, is now the regulator’s starting point. Advisory tools belong in it. More on building one: how to find where AI is already used on a site.
S-OIL announces AI in its process safety management system for the Shaheen plant
22 October 2026.
S-OIL, the South Korean refiner, announced that it has completed the rollout of its integrated manufacturing operations management system, which brings production, facilities, maintenance, inspection and safety data onto one platform. The latest phase links its master data and its reliability and integrity management system. It also adds eight modules to its process safety system, among them management of change, incident management and process hazard analysis, alongside the four existing modules built around digital work permits. The job safety analysis module uses generative AI to suggest the hazards of each task, drawing on about 2,800 standard work risk assessments, past incidents and injury and fatality risk factors. The system is being extended to the Shaheen petrochemical plant in Ulsan before it starts commercial operation. This is the company’s announcement as reported in the press, in machine translation; the report mentions no published safety validation of the AI module (The Herald Business).
What it means for operators: once AI proposes the hazards in a job safety analysis, it shapes control of work at the point of the task. The site has to decide who checks the suggestions, how a missing hazard is caught and how the hazard database behind it is kept up to date, and to treat the introduction as a change to the management system. More: AI and control of work.
Havtil audit findings on new technology at Vår Energi and Aker BP
22 October 2026.
Two recent audits by Havtil show what inspectors now look for when operators introduce new technology. Neither set of findings is about the technology itself: each is about whether the organisation assessed the change and its risks before using it.
- Vår Energi, automated drilling. Audit in December 2025, order on 5 February 2026: three non-conformities, in qualification of technology, analyses and non-conformity handling, which Havtil judged “serious breaches of the regulations”. The company had no technology qualification process and had used its management of change process instead, and its acceptance criteria were general objectives without concrete human, organisational and technical criteria. Havtil ordered criteria for the development, testing and use of the technology before the systems are taken into use, and a binding plan within two weeks (Havtil).
- Aker BP, Yggdrasil. Audit in June 2026 of risk management and qualification of new technology such as AI, published on 22 October 2026: two non-conformities, in technology qualification, and in risk management and risk understanding linked to the planned use of robots. Aker BP is to report how it will deal with them by 15 November 2026 (Havtil).
What a site should be able to show an inspector is set out in the regulatory audit readiness checklist and in what to do after a regulator letter or order on AI.
British Columbia mines: repeated incidents with autonomous haul trucks
18 October 2026 (weekly reports since 4 January 2026).
The British Columbia mines regulator publishes a short weekly summary of each reportable incident, without naming the mine. In 31 of its weekly reports from January to 18 October 2026, 46 entries involve autonomous haul trucks. Two patterns recur. Trucks slid on ramps and ended in the oncoming lane, some at 35 to 47 km/h, with berm contacts and a truck stopping about 3 metres from a light vehicle at an intersection. Vehicles and people entered active autonomous zones: several entries record an unauthorised vehicle entering “without the required layers of protection”, and others a contractor’s vehicle, a mechanic doing repairs and workers inspecting a cable. Fires and tyre failures on autonomous trucks also recur. The entries are short summaries, not investigation reports, and give no causes (weekly reports of 1 March, 28 June and 18 October 2026).
What it means for operators: at autonomy level 4 the hazards move to the edges of the autonomous zone: who may enter it, under what isolation, and the limits on ramp and road conditions. Those sit in control of work and operating limits, not in the vehicle software. More on both: AI and control of work and AI and operating limits.
NORCE research: keeping people in control of autonomous drilling
31 August 2026.
A paper by NORCE researchers, summarised by Havtil, works through three cases: an autonomous drilling system acting in real time on uncertain data, an AI-based managed pressure drilling system supervised by an operator, and planning tools based on language models. As summarised by Havtil, more autonomy moves the operator from active control to monitoring, assessment and intervention, and makes errors, faulty sensor data and interactions between systems harder to see. Operators and supervisors need different information: uncertainty and actions for the first, system limits for the second. A “safe mode” function can hold the system in a temporary safe state while control is handed back to people, and the paper asks how conflicting assessments from different suppliers’ systems can be detected and flagged automatically (Havtil).
What it means for operators: handover to manual control is a designed function with its own limits, procedures and training, not an assumption. More on the competence side: operator competence when AI runs the plant.
UK HSE sets out how health and safety law applies to AI
12 June 2026.
The Health and Safety Executive states that the goal-setting Health and Safety at Work Act already covers AI, without AI-specific rules. Duty holders must risk-assess uses of AI that affect health and safety, apply controls so far as is reasonably practicable, and include cyber threats. The statement refers to major accident hazard sites and links to its COMAH guidance. Planned work for 2026/27 includes new guidance on applying health and safety law to AI, a consultation on incident reporting (RIDDOR) and an innovation hub. Research under way covers safe use of AI in process safety applications, supported by the AI Security Institute, and cognitive offloading in AI-assisted safety decisions in major hazard and critical infrastructure industries (HSE).
What it means for operators: in Great Britain, an AI use that affects safety needs a risk assessment now, under existing law, before any new guidance arrives. More on what regulators expect: after a regulator letter or order on AI.
Borouge reports an AI autonomous operations pilot at Ruwais
9 June 2026 (Borouge announcement 7 January 2026).
Borouge announced that it had completed, in collaboration with Honeywell, a proof of concept of AI-powered autonomous operations in a live production environment at its Ruwais petrochemical facility, after trials in 2025, with the aim of an AI-driven control room. The company reported potential gains in efficiency, downtime and operating cost. The unit concerned, the control mode (advisory or closed loop) and the role left to operators are not stated (Borouge). In a vendor announcement in June, Honeywell launched the control room software behind it, described as making recommendations and automated decisions through AI-powered agents that act on behalf of the operator (Honeywell).
What it means for operators: neither announcement says anything published on oversight, override, management of change or the safety case. An operator considering the same step should treat those as the first questions. More: AI in process control and the DCS.
Havtil orders to Equinor: control room alarms at Johan Sverdrup and qualification of Novos
22 May 2026 (earlier order 19 December 2025).
Havtil audited the Johan Sverdrup central control room and driller’s cabin in January 2026, looking at human factors and the extra load from a second process platform controlled from the same room. It found five non-conformities: follow-up of the alarm system, human-machine interfaces, working environment analyses, non-conformity handling, and follow-up and evaluation of the automated One Button Startup. It judged the alarm findings a serious breach: too many new alarms, in normal and critical situations, for operators to perceive and handle in time. On 12 May it ordered measures to reduce the alarm rate by 31 August 2027, with compensating measures within eight weeks of its notification (Havtil). In December 2025 Havtil had ordered Equinor to qualify Novos, an automated drilling control (ADC) technology that Equinor kept at technology readiness level 4, before using it on further facilities, by 1 June 2026 (Havtil).
What it means for operators: when automation widens what one control room runs, alarm load and interfaces have to be re-assessed with it, and a technology in use must be qualified to the level of its use. More: is your organisation ready for a regulatory audit?
UK nuclear regulators test AI in a regulatory sandbox
1 May 2026.
The Office for Nuclear Regulation led a seven-month sandbox with the Environment Agency, HSE and the Defence Nuclear Safety Regulator, working with industry partners including Rolls-Royce Submarines, NNB Generation Company, Sellafield and Nuclear Restoration Services on two applications of machine learning to images and video for monitoring and inspection. The findings: the UK’s outcome-focused framework can accommodate AI, safety depends on design and implementation, and a wider range of uses can be pursued safely than industry first proposed. The lessons fall in three areas: the skills to develop and assess AI, how to give assurance for it, and how it fits into existing nuclear safety cases. The report is not a formal regulatory position (ONR).
What it means for operators: the same three questions apply in any major hazard industry: who can assess the AI, what evidence assures it, and where it sits in the safety case. More: autonomous operations beyond oil and gas.
Chemical companies place themselves at autonomy levels 1 to 2
9 February 2026 (NAMUR meeting November 2025).
At the NAMUR meeting of process automation users in November 2025, most companies placed themselves between levels 1 and 2 of the autonomous operation scale, according to a trade press report. Olaf Abel of BASF is quoted saying that basic automation is often insufficient, with high investment needed and savings not always visible (POWTECH TECHNOPHARM). A Schneider Electric survey of 400 executives, reported in April 2026, put energy and chemical companies together at 3.52 out of 5 on the ARC maturity model (Chemical Processing).
What it means for operators: self-assessed levels differ widely depending on who asks and how. A level is only useful if it is stated unit by unit against a defined scale. More: the autonomous operation maturity levels.