OCPredictive maintenance is often sold as a technology problem. In practice, the algorithm is one of the last pieces that matters.
An AI model can detect an unusual vibration, temperature drift or pressure pattern. It can flag a weak signal before a person would have noticed it. What it cannot do is repair a weak maintenance strategy, poor equipment hierarchy, unreliable work-order history, missing failure modes or unclear decision rights.
If the operating model is not ready to act on the signal, prediction becomes another alert.
01Start with the maintenance strategy, not the model
The first question is not, “Which predictive platform should we buy?” It is, “Which failures are we trying to prevent, and why do they matter?”
Criticality, failure consequence, maintenance task, detection method and response window need to connect. A sensor on non-critical equipment may create more noise than value. A critical asset with the wrong failure mode mapped may create false confidence.
A useful predictive-maintenance case begins with an asset hierarchy and a clear view of failure mechanisms. From there, the organisation can decide what should be run to failure, what should be maintained on a fixed interval, what should be condition-based and where prediction is genuinely justified.
02Data quality is an operating discipline
Industrial AI depends on data, but “more data” is not the same as better evidence. Sensor history needs to be correctly tagged. Equipment records need to identify the same asset across operations, maintenance and engineering systems. Work orders need failure codes that mean something. Maintenance notes need enough consistency for patterns to be interpreted.
TotalEnergies made the same point explicitly in June 2026: data is the key enabler for scaling AI. The company said nearly 3,000 pieces of equipment were already being monitored, with AI analysing equipment data to detect early warning signs and help teams plan maintenance before failures occur. Its ambition is to extend that approach to tens of thousands more assets.
That is significant, but the important word is not AI. It is monitored. The infrastructure, instrumentation, data model and operating routines have to exist before an algorithm can produce a useful maintenance decision.
03A prediction only creates value when it changes work
An anomaly score has no commercial value on its own. Someone must decide whether to inspect, defer, isolate, repair, order parts or change the operating envelope. The work then has to move through planning, permits, spares, contractors, shutdown windows and operational priorities.
This is where many predictive initiatives stall. The analytics team proves that a model can detect something. The maintenance organisation still runs on the old cadence, the planner does not trust the signal, the work-order process has no route for predictive interventions, or the operating team cannot see the rationale behind the recommendation.
The question therefore becomes: what decision does the signal trigger, who owns that decision, and how quickly can the organisation act?
04Domain knowledge remains part of the system
ExxonMobil describes its approach to AI in similar terms: sensor data becomes useful when it is combined with deep industry understanding and domain knowledge. That matters because industrial assets do not fail in tidy statistical isolation. Operating context, maintenance history, process conditions, recent interventions and known design weaknesses all affect whether a signal is meaningful.
The best predictive-maintenance systems therefore do not remove engineers, operators and maintainers from the loop. They make their judgement more timely and better informed.
The objective is not to replace maintenance expertise. It is to focus it earlier, where intervention costs less and disruption is lower.
05The corrective-maintenance baseline tells you whether you are ready
In an offshore gas operating-model study I led, around 80% of maintenance hours were corrective. That is not simply a maintenance statistic. It is a description of the operating model.
Corrective work consumes planning capacity, creates unpredictable demand for people and spares, complicates access and logistics, and increases the number of interventions required on the facility. The work therefore had to go beyond selecting technology. We mapped the manual interventions, identified instrumentation gaps, redesigned the maintenance approach and built a roadmap towards more predictive and preventive work.
That sequence matters. Installing AI on top of an 80% corrective environment without changing the underlying maintenance system would have produced better warnings inside the same reactive organisation.
06Five foundations to put in place before predictive AI
- Asset criticality: agree which equipment matters most and what the consequence of failure actually is.
- Failure modes: define the mechanisms worth detecting and the lead time required for action.
- Trusted data: align tags, equipment hierarchy, condition data and maintenance history across systems.
- Work-process integration: specify how an alert becomes an inspection, work order, materials decision or operating intervention.
- Decision rights and KPIs: make ownership clear and measure avoided failure, intervention quality and maintenance mix, not the number of alerts generated.
07Use AI after the foundations are visible
The current industrial direction is clear. TotalEnergies is scaling AI-enabled monitoring across industrial equipment, while its MethaneLive programme combines 13,000 sensors, real-time data, advanced algorithms and specialist teams to identify anomalies and target corrective maintenance. ExxonMobil is using AI to connect sensor data across its operations and improve decisions through domain knowledge.
These examples do not suggest that technology is unimportant. They show the opposite: advanced analytics becomes more useful when the surrounding foundations are strong enough to carry it.
The Oclas method is deliberately less glamorous than starting with software:
Map documented process → validate beside people doing the work → quantify differences → separate BAU fixes from system change → embed through KPIs, training and coaching.
For predictive maintenance, that means understanding how the asset is actually maintained before designing how AI should support it.
Prediction is not the maintenance strategy. It is a capability inside one.
Sources: TotalEnergies, Data & AI in Support of the Transition Strategy, 18 June 2026; TotalEnergies, MethaneLive, 18 June 2026; ExxonMobil, AI as a catalyst for innovation, 14 April 2026.