FrameworkOil & Gas

From Reservoir to Refinery: Where AI Earns Its Keep

A value-versus-readiness framework for AI in oil and gas, from subsurface to trading, and the four failure modes that stall pilots before they reach the plant.

Arclight Research5 min read

Fig. I · Oil & Gas

In brief

  1. IPlot every AI use case on value at stake against readiness, then fund the readiness gaps instead of launching more pilots.
  2. IIReadiness is four tests: usable data, outputs that reach the decision in time, a path through management of change, and an owner.
  3. IIIGood models fail on data quality, integration, safety cases and drift. Budget monitoring and retraining as an operating cost.

Oil and gas companies were using mathematical decision tools long before anyone called them AI. Reservoir simulation, refinery linear programs and advanced process control have shaped capital and operating decisions for decades. That history is an advantage and also a trap, because it invites the assumption that a new model will slot into operations the way the last optimizer did.

For most executive teams, the open question is no longer whether AI works in the field. It is which applications deserve capital, and why so many promising pilots stall before they reach a control room.

The value-readiness lens

Plot each candidate application on two axes. The first is value at stake, measured in the asset’s own economics: barrels, uptime, avoided emissions, safety exposure or working capital. The second is readiness, the axis most often misjudged, and it comes down to four tests that an application must pass together:

  • Data. The signal is measured at the right frequency, labeled with real outcomes and accessible outside the historian.
  • Integration. The output reaches the person or system that acts on it within the decision window, whether that is seconds on a compressor or months before a turnaround.
  • Safety. The application can be taken through management of change and the facility’s safety case without redesigning either.
  • Ownership. An operations leader owns the model’s performance after go-live, long after the project team has moved on.

If a model cannot survive management of change, it was never ready for the plant.

The resulting grid is simple. Scale what is high on both axes, and where value is high but readiness is low, fund the readiness work as a deliberate program. Buy, rather than build, applications that are ready but modest in value. Stop the rest.

Upstream: subsurface, drilling and production

In the subsurface, computer vision speeds fault and horizon interpretation on seismic volumes, and machine-learned proxies for reservoir simulators make it practical to run far more history-matching and uncertainty cases. The value is high, but readiness is often overstated. Labeled data is sparse, geology does not transfer cleanly between basins, and a proxy trained on one range of conditions can fail quietly outside it. Physics-informed models narrow that risk by constraining predictions to what the reservoir can physically do, but they do not remove it.

Reserves add a governance test. The SEC’s definition of reliable technology, which expressly includes computational methods, requires field testing and results that are reasonably certain, consistent and repeatable. If a model’s output supports booked reserves, the company must be able to show that standard is met, and its reserves auditors will ask.

Drilling and production sit closer to the top of both axes. Rigs stream dense real-time data, and models for rate of penetration, stuck-pipe risk and kick detection can act inside the decision window. In production, anomaly detection on artificial lift, classification of rod-pump dynamometer cards and virtual flow metering move surveillance from reviewing every well to managing by exception.

The constraints here are commercial as much as technical. On many rigs the drilling contractor controls the control-system data, and the contract decides who may use it. Production models also depend on well tests and allocation factors, which can be less reliable than the sensors feeding the model.

Methane and midstream

Methane detection has moved from periodic camera surveys to layered monitoring: continuous sensors on site, aerial surveys above it and satellites above that. Machine learning matters at two points, attributing a plume to its source and reconciling measured emissions with bottom-up inventories, which is the direction of measurement-based reporting under frameworks such as OGMP 2.0. The value is regulatory, commercial and reputational, since gas and LNG buyers increasingly ask for measured intensity.

The failure mode is false positives. Consider an operator that installs continuous monitors across a basin and dispatches a crew for every alert. Within a season the crews learn which sensors to discount, and the one real leak waits in the same queue as the false ones.

Midstream applications follow the same pattern. Models that prioritize in-line inspection anomalies, predict compressor failures and screen satellite imagery for right-of-way encroachment are ready to deploy where the data is clean. In the United States, pipeline control rooms already operate under regulated alarm management, so a model that adds noise to the console is a compliance problem as well as a nuisance.

Downstream, turnarounds and trading

Refineries have run linear programs for crude selection and advanced process control for decades, so AI works at the margins: soft sensors that infer product qualities between lab samples, fouling prediction for heat exchangers and predictive maintenance on rotating equipment. These rank high on readiness where historian data is clean, and their value concentrates in avoided unplanned shutdowns and slowdowns.

Turnarounds are a larger prize at lower readiness. Scope is assembled from inspection history, risk-based inspection results, work orders and vendor documents, much of it unstructured. Document intelligence and retrieval-augmented generation can build draft work packs and surface scope risk earlier, provided the output flows into the planning system and is checked by the people who will sign off the work.

In trading and supply, optimization of blending, vessel scheduling and nominations is proven ground; price forecasting is not. Language models are useful for contract review and communications surveillance. They are dangerous when allowed to summarize a position or a counterparty’s obligations without a trader checking the source.

Four ways good models fail

Data quality

Historians compress data for storage, tags are mislabeled and instruments drift out of calibration. A model trained on that record learns its errors. The fix is unglamorous work on tag governance and instrument maintenance, funded before the model.

OT/IT integration

Control networks are segmented for good reasons. A model that needs data to cross those boundaries must be designed with the cybersecurity team from the start, and anything that writes back to a control system belongs in a separate and far more demanding category.

Safety cases

An AI recommendation that changes how operators run a unit is a change to the operation. It goes through management of change, is assessed against the safety case or process safety program, and stays out of safety instrumented functions. If a model cannot survive management of change, it was never ready for the plant.

Model drift

Reservoirs deplete, crude slates change, equipment is replaced and seasons turn. Models degrade silently unless someone monitors them against outcomes and owns retraining, which is why that cost belongs in the operating budget rather than the project budget.

Questions for your next leadership meeting

  1. Which of our AI pilots pass all four readiness tests today, and what specifically blocks the rest?
  2. For each model in production, who owns its performance, and when was it last checked against actual outcomes?
  3. Does any model output feed booked reserves, emissions reports or safety-related decisions, and could we defend it to an auditor or regulator?
  4. Who controls the data our operations generate: us, our drilling contractors or our equipment vendors?
  5. What share of our AI budget goes to data, integration and monitoring, and is it enough to keep what we have already built working?

ARCLIGHT · RESEARCH

This analysis is general in nature and does not constitute investment, legal or engineering advice. To discuss how it applies to your organization, request a private briefing.

Begin

Apply this analysis to your own assets.

Tell us about your assets, your board and the decision in front of you. A principal replies within one business day.