AI for Oil and Gas
Oil and gas has the densest sensor coverage in heavy industry, the longest-lived assets, and a safety case that constrains every optimisation before it constrains any model.
The engineering here is rarely the hard part. The hard part is that every recommendation has to be defensible inside a safety case written by people who assume things will go wrong, and rightly.
AI for oil and gas covers production optimisation and artificial lift, equipment condition monitoring on rotating and static plant, subsurface and seismic data analytics, pipeline and asset integrity, HSE and incident analytics, and operations decision support.
Optimisation inside a safety case
Production optimisation asks a model to move setpoints toward better output. The safe operating envelope exists because a HAZOP identified what happens outside it. Those two facts have to be reconciled explicitly rather than by hoping the optimiser behaves.
- The envelope is a hard constraint, never a penalty. An optimiser will find the edges of whatever window it is given, because that is what optimisation does.
- Constraints come from process safety, not from the modelling team. They are inputs to the build, agreed and signed, not parameters to be tuned.
- Where a recommendation approaches a limit, it should say so. Not approach it and report success afterwards.
- Advisory to the control room, not into the control loop. A recommendation an operator accepts is a different regulatory and safety proposition from an automated action.
- Log every recommendation and every override. Both are evidence, and the override pattern tells you what the model is missing.
Operator overrides are the most informative data you will collect
A production optimisation system that is never overridden is being ignored; one overridden constantly is missing something operators can see and the model cannot. Track override rate and the reason, and treat a consistent override pattern as a missing feature rather than a training problem. In several engagements the override analysis has been more valuable than the optimiser, because it surfaced operational knowledge that existed only in individual heads.
Condition monitoring, and the failure-example problem
Rotating equipment — compressors, pumps, turbines — is well instrumented and expensive to lose. It is also maintained carefully enough that catastrophic failures are rare, which removes the training data that failure prediction requires.
Monitor deviation from known-good, not failure
Learning the normal operating signature of a machine and detecting departure from it needs no failure examples at all, works from the first month, and catches degradation trends. It is a weaker claim than prediction and it is the claim the data supports.
Sampling rate decides what is possible
Vibration analysis needs high-frequency waveform data. A historian recording one averaged value per minute has discarded the diagnostic content permanently, and no model recovers it. Check the resolution before scoping anything.
Operating context is a first-class variable
The same machine behaves differently at different loads, speeds, ambient conditions and product compositions. A model without operating context will flag normal load changes as anomalies and be switched off within a fortnight.
Maintenance records are free text and worth structuring
What was actually found and fixed is the label that makes everything else better, and it usually sits in unstructured work order notes. See text classification.
Subsurface, integrity and HSE
| Application | Maturity | Note |
|---|---|---|
| Seismic interpretation support | Good | Assists interpreters; the geological judgement stays human |
| Well performance and artificial lift | Strong | Well-instrumented, measurable, clear economic return |
| Pipeline integrity from inspection data | Strong | Large image and signal volumes; review capacity is the constraint |
| Corrosion and erosion prediction | Moderate | Physics-informed approaches beat pure data-driven here |
| Drilling optimisation and NPT reduction | Moderate | Real value; requires rig data quality that varies enormously |
| HSE incident and near-miss analytics | Good | Years of free-text reports nobody aggregates |
| Autonomous process control | Not ready | Safety component territory. Different programme entirely. |
Incident reports are an unmined safety asset
Near-miss and incident reports accumulate for years, are read individually at the time of filing, and are almost never analysed in aggregate. Classified and grouped, they show recurring patterns by asset, shift, contractor, procedure and task type — which is exactly what HSE functions want and rarely have the capacity to produce. It is a contained piece of text work with a safety return rather than a production one, and it is usually the easiest project in the business to get approved.
How an engagement runs
The safety envelope is established as a hard constraint before any optimisation is designed.
Data and constraint assessment
Historian resolution and tag reliability, maintenance record quality, and the operating envelope stated by process safety as inviolable.
Data foundation
Historian, maintenance, operating context and inspection data joined; sampling adequacy verified before modelling.
Build
Production optimisation with constraints encoded, or condition monitoring against known-good signatures with operating context modelled.
Advisory trial
Recommendations to the control room, with acceptance and override rates tracked from day one.
Operation
Constraint adherence audited, override patterns reviewed as missing features, drift monitored as plant condition changes.
What you receive
Better production and earlier warning, inside a safety case that still holds.
Safety constraint set
The operating envelope from process safety, encoded as hard constraints rather than penalties.
Data adequacy assessment
Whether your historian resolution supports the intended analysis, stated before modelling.
Production optimisation
Advisory to the control room, with recommendations that decline rather than approach limits.
Condition monitoring
Against known-good signatures with operating context, needing no failure history.
Structured maintenance records
What was found and fixed, extracted from work order free text.
HSE incident analytics
Recurring patterns by asset, shift, contractor, procedure and task.
Is this the right starting point?
Worth being direct. There are situations in oil and gas where custom AI work is the wrong spend, and those are listed rather than buried.
Worth doing if
- Production optimisation is done by experienced operators without decision support.
- Rotating equipment is well instrumented and monitored only against fixed alarm limits.
- Maintenance and HSE records contain years of free text nobody has analysed.
- Pipeline or asset inspection generates more data than review capacity can absorb.
- You want the safety-component classification settled before building anything.
Do something else if
- Historian sampling is too coarse for the intended analysis and cannot be changed.
- You want an optimiser with authority over the safety envelope. We will not build that.
- You want failure prediction on assets with almost no failure history.
- The constraint is reservoir behaviour or market price, which no model changes.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
Will an optimiser push us outside safe operating limits?
Not if the limits are encoded as hard constraints set by process safety, which is how we build them. An optimiser finds the edges of whatever window it is given — that is what optimisation is — so the window must come from your safety case and be treated as inviolable. Where a recommendation would approach a limit, the system should say so rather than approach it and report the result.
Why does our condition monitoring generate so many false alarms?
Usually because operating context is missing. The same machine behaves differently at different loads, speeds, ambient conditions and product compositions, so a model without that context flags normal load changes as anomalies and gets switched off within a fortnight. Model context as a first-class variable rather than trying to filter its effects out afterwards.
Can you predict equipment failures?
Rarely, and we would rather build what works. Well-maintained rotating equipment fails infrequently, so the failure examples that prediction requires do not exist — you might have decades of operation and a handful of relevant events. Condition monitoring against a known-good signature needs no failure history, works from the first month, and catches degradation trends. It is the honest version of the same value.
What is the easiest project to get approved?
HSE incident and near-miss analytics, usually. Years of free-text reports sit unread in aggregate, and classifying them reveals recurring patterns by asset, shift, contractor and procedure. It touches neither production nor the safety envelope, the return is a safety return rather than a production one, and it clears internal approval far faster than anything near the process.
Is our production optimisation caught by the AI Act?
Advisory optimisation that recommends setpoints to an operator is generally outside the critical infrastructure high-risk category, which turns on whether a system is a safety component whose failure could cause physical damage or harm. A system with authority to act on plant is a different matter. The line is finer than it looks, regulators are expected to read it broadly, and the reasoning should be documented whichever way you conclude.
Related verticals
Organisations in oil and gas usually share data, buyers or regulators with these. All fourteen are listed on the Energy & Utilities page.
AI for Power Generation and Utilities
Demand forecasting, network asset health, inspection analytics and outage response — advisory, not actuating.
Read more →AI for Renewable Energy
Generation forecasting, turbine and inverter health, and curtailment — where measurement beats modelling.
Read more →AI for Water and Wastewater Utilities
Leakage detection, treatment optimisation and overflow prediction — where public accountability is the constraint.
Read more →Tell us what the problem looks like.
Thirty minutes, no charge, no deck. We will tell you whether this is an AI problem, a data problem, or a process problem — and we will say when the honest answer is to buy something rather than build it.