An AI tool does not need permission to operate a machine to change the risk around it. If it persuades a maintenance team to postpone an inspection, the decision can affect equipment, people and production. The question for management is not simply whether a person approves the recommendation. It is whether that person has enough information, time and authority to challenge it.
A Norwegian petroleum industry survey makes the distinction concrete. In October 2026, the Norwegian Ocean Industry Authority, Havtil, reported that some operators described using AI for well planning, maintenance and operational decision support while saying they had no AI systems with implications for safety. Havtil cautioned that advice and priorities can influence safety even when AI does not directly control equipment.
This matters beyond petroleum. A Norwegian manufacturer, food producer or transport company may use an AI-supported maintenance system without thinking of it as safety technology. The right response is not to stop every useful tool. It is to identify the decisions that could affect people or essential operations and establish how those recommendations are tested and challenged.
What Havtil actually found
In its survey published on 21 October, Havtil asked nine operators on the Norwegian continental shelf twelve questions about their use and governance of AI. All nine replied in writing. The underlying report says the responses were not verified through inspections or interviews. This is a small, sector-specific survey, not a measure of Norwegian business as a whole.
More than half reported that they did not have, or could not document, a complete inventory of AI systems under development and in use. Some had established registers and responsibilities; others had limited visibility, including of AI embedded in suppliers' products. Havtil also found that companies described different degrees of worker involvement and supplier requirements. The survey points to gaps in oversight, not proof that an AI system caused an accident.
Havtil's interpretation of the petroleum health, safety and environment rules applies to the activities it regulates. It should not be treated as an identical legal requirement for every Norwegian AS. Companies in other sectors need to assess the rules that apply to their own activities.
A recommendation changes a decision
Imagine a Norwegian food producer with a cooling pump. The example is hypothetical. A supplier's system recommends delaying an inspection because recent sensor readings look normal. A maintenance manager accepts. Later, a fault develops that the sensors did not capture, interrupting production and potentially affecting food safety.
The example is not an assertion that this happened or that AI is less reliable than manual scheduling. A useful system might also identify deterioration earlier than the previous routine. The point is that the real decision includes the sensor, the data, the recommendation, the operator and the consequences of delay. If the system misses a warning, a person must know what independent evidence to check and when to stop relying on the recommendation.
Before moving an inspection, ask the supplier and the people who maintain the equipment:
- What information supports the recommendation? Identify the sensors and records used, how old they are and what the system cannot see.
- What happens when information is missing or wrong? Test a failed sensor, a stale reading and an unusual operating condition. An output that looks normal is not proof the inputs are sound.
- Who can disagree and what happens next? Give the responsible employee authority to retain the inspection or escalate it. Record the reason for a material override and follow up on the result.
- What remains independent? Confirm which inspections, alarms and established safety measures continue regardless of the recommendation. Do not assume that two alerts are independent if they rely on the same sensor.
These are practical questions, not a substitute for an engineering assessment. For critical equipment, qualified staff must define the safe operating limits and test plan for the specific installation.
A person in the loop needs a real choice
A name beside an approval is not enough. The person making the decision needs the original evidence, a clear indication of missing data and a workable route to ask for help. A team pressed to approve hundreds of recommendations may not have a meaningful opportunity to investigate each one. Management must decide which cases need deeper review, provide the time to do it and check whether the arrangement works in practice.
A separate Havtil audit of planned AI and robotics projects at Aker BP, published on 22 October, illustrates the question. The authority reported two nonconformities concerning technology qualification and risk management for planned robot use. It said it could not verify that human interaction with AI had been systematically assessed under different operating and failure conditions. The audit concerned projects in development, not an AI-caused accident. Havtil asked the company to explain how it would address the findings by 15 November 2026.
The lesson for other companies is to test the decision, not just the model. Present a realistic case with incomplete or conflicting evidence. Can the employee identify the uncertainty, challenge the suggested action and keep the equipment safe? Listen to maintenance staff and employee representatives when designing that test. Their knowledge of actual work and abnormal conditions matters alongside the supplier's technical results.
What the supplier should demonstrate
A supplier may know the model while the customer knows how the equipment is used. Neither has the full picture alone. Ask for evidence of performance in conditions relevant to your operation, including missed faults and unnecessary alerts. Clarify what is measured, how changes to the model or sensors are communicated, who investigates failures and what service continues if the tool is unavailable. A single accuracy score does not answer these questions.
Testing has a cost. Keeping some inspections that a system recommends delaying takes time and may reduce short-term savings. Yet accepting every recommendation without checking can hide a deteriorating system until the cost is larger. Start with equipment where a wrong decision would matter most, compare recommendations with actual inspection findings, and adjust the review as evidence improves. In a small company this may mean one critical machine and a meeting with the existing maintenance supplier, not a new technology department.
Management acts, the board checks the direction
Management should identify the uses of AI that influence maintenance or other important operational decisions, appoint an owner for each, involve the people doing the work, and agree on tests and an escalation route with suppliers. The record should show what was tested, what remains uncertain, and when the next review is due.
Where operational safety is a material risk, the board should ask whether management knows which decisions are affected, whether staff can challenge recommendations, and what evidence shows the arrangement works. It should receive material exceptions and unresolved risks, rather than approve individual maintenance schedules.
The same distinction applies when moving an AI pilot into ordinary operations: successful trials do not remove the need for an owner, monitoring and a way to recover when a system fails. For decisions with financial rather than physical consequences, the guide to challenging AI advice explores a different but related problem.
A good AI recommendation can make maintenance more effective. The test is whether the business can explain what it will do when the recommendation is incomplete, wrong or impossible to check.
Sources were checked on 1 November 2026. The food producer and pump are hypothetical. The practical steps are the author's recommendations; legal obligations depend on the business and its activities. Illustration generated with AI.