AI in Oil and Gas: Where It Works, and Where It Has to Stop

How AI is used in oil and gas, from drilling to refining: where it fits on the equipment, what CISA and the standards say, and whether it will replace jobs.

AI in oil and gas is real, and it is narrower than the conference keynotes suggest. In the documented cases we could trace to an operator, AI does one of two things. It advises an engineer, who decides. Or it controls one bounded setpoint, such as the auto driller's rate of penetration or a well's gas lift injection rate, inside limits a person set. We found no published case of AI inside a safety instrumented system, and the joint guidance from US and allied cyber agencies says AI such as large language models almost certainly should not be used to make safety decisions in OT.

This page is for two groups: operators deciding where AI belongs on their assets, and the automation, controls and software companies that sell to them. It covers what is deployed, where it fits on real equipment, what the regulators and standards bodies say, and what it means for jobs.

How is AI being used in oil and gas?

These are documented deployments, each traced to an operator or vendor announcement or a technical paper. Most are reported by the operator or the vendor, so treat the performance figures as their claims, not independent measurements.

Stage	Who	What the AI does	Who acts on it (our reading)
Subsurface	Shell, with SparkCognition	Generative AI builds deepwater seismic images from far fewer shots, as little as 1% in field trials^{1, 2}	Advisory
Drilling	Equinor, with SLB, Peregrino C, Brazil	99% of a 2.6 km section drilled in autonomous control mode; 60% faster rate of penetration over five wells, against a baseline not disclosed³	Closed loop on drilling parameters
Drilling	Equinor, with Halliburton and Sekal, Norwegian shelf	Automated on-bottom drilling that keeps parameters within modeled safe limits⁴	Closed loop on drilling parameters
Drilling	Nabors and Corva, for an unnamed Delaware Basin operator	Closed-loop control of the auto driller; 36% average gain in rate of penetration over four wells⁵	Closed loop on auto driller setpoints
Well pad	ExxonMobil, Bakken	Machine learning gas lift optimization deployed on 200+ wells; about 7% average uplift on the first 50 wells optimized^{6, 7}	Not stated; results feed monthly engineer reviews
Well pad	Permian operator, 1,300+ wells (SPE 219553)	Well by well gas lift optimization through remote injection-rate setpoints⁸	Closed loop on injection rate
Offshore and plants	Shell, with C3 AI	Predictive maintenance on 10,000+ pieces of equipment: control valves, pumps, compressors⁹	Advisory
Offshore	Equinor, Johan Sverdrup	Digital twin of the platform used by operators on tablets; Equinor credits its digital tools with more than NOK 2 billion in year one¹⁰	Decision support
Gas processing	ADNOC, Bab field and Taweelah compression plant	Process optimization and predictive maintenance; the pilot showed potential to cut unplanned shutdowns by 50%¹¹	Not stated
Gas processing	Aramco, Fadhili Gas Plant	Reinforcement learning on acid gas removal; Aramco reports 10% to 15% less amine and steam use¹²	Not stated
Refining	TotalEnergies Port Arthur, with Honeywell	In a pilot with preliminary results, an AI assistant in the control room forecast five events, on average about 12 minutes before the alarm¹³	Advisory

One we left out of the table: Flyscan, backed by Kinder Morgan, Enbridge and Marathon Petroleum, sells AI leak detection from patrol aircraft, but we found no operator report of it in use¹⁴.

Two patterns stand out. First, the closed-loop cases are all narrow: a drilling parameter or an injection rate, with the envelope set by people and the safety systems untouched. Second, the biggest numbers come from fleets, not single assets. Shell's predictive maintenance program runs nearly 11,000 models across more than 3 million sensors, by C3 AI's count⁹. That is the scale at which the question of who owns each recommendation stops being theoretical.

Where AI fits on the equipment

Every piece of operational technology (OT) on an oil and gas asset already produces a stream of measurements: the PLCs, drives, sensors and safety systems that run the machinery, as opposed to the IT that runs the office. That is what makes each one a candidate for AI. The table is our read of where analytics add the most across the value chain, and who should still own the decision. It is a map of opportunity, not a list of what any operator runs today.

Device	Where	What AI could add	Who should decide
Drilling control system	Rigs	Tune rate of penetration; warn of stuck pipe from torque and drag patterns	Driller, inside limits set by the well plan
Mud pit and flow sensors	Rigs	Tell a real kick from a routine pit transfer, sooner	Driller. Shutting in the well stays a human, hydraulic action
Blowout preventer control system	Rigs	Trend element wear and test results to plan maintenance	Well control engineers. AI stays out of the actuation path
Rod pump and gas lift controllers	Well pads	Set pump speed or injection rate well by well	Production engineer, within set limits
Compressors and rotating equipment	Platforms, FPSOs, compressor stations, LNG	Predict faults and estimate remaining useful life	Maintenance supervisor
Separators and process control (DCS)	Platforms, gas plants, refineries	Recommend setpoints; forecast upsets before the alarm	Control room operator
Pipeline SCADA and leak detection	Midstream	Separate real leaks from instrument noise and pressure transients	Pipeline controller
Fire and gas detectors	Every asset	Catch detectors drifting out of calibration; filter nuisance alarms	Installation or plant manager. Trips stay in the certified system
Safety instrumented system	Every asset	Watch its health and help plan proof tests. Nothing more	Safety engineer. AI never fires or holds off a trip
Network gateways between OT and IT	Every asset	Flag unusual traffic on the control network	OT security team

We map these device by device on interactive models you can take apart in a browser. The jack-up rig has eleven OT devices mapped. The offshore oil platform walks through a fixed production platform. The oil pump jack maps eight places on an onshore well pad. The exhibits page has every model, including the semisubmersible. An FPSO and an LNG plant are next.

Three rules for AI on operating assets

The pattern in the right hand column is the point: the higher the consequence, the further AI should sit from the actuator. We hold any of these systems to three rules.

Advise, then decide. The model recommends, a named person decides, and the record keeps both: what the model saw, what it said, and what was done.
Stay out of the safety path. Well control, emergency shutdown and fire and gas trips stay in certified, independent systems. AI can watch them and help plan their maintenance. It should not be the thing that fires them or holds them off.
Know where the model learned. A model trained on one basin, one rig type or one plant is not evidence for another. Validate on the asset class, and say so in writing before anyone relies on it.

What the regulators and standards bodies say

The clearest statement so far is joint guidance published on December 3, 2025 by CISA, the NSA, the FBI and cyber agencies from Australia, Canada, Germany, the Netherlands, New Zealand and the UK: Principles for the Secure Integration of Artificial Intelligence in Operational Technology^{15, 16}. Three lines from it carry most of the weight:

"AI such as LLMs almost certainly should not be used to make safety decisions for OT environments."
"Ultimately, humans are responsible for functional safety."
"Ensure there are failsafe mechanisms that revert to traditional automation or manual for any AI-enabled system processes."

The same guidance favors architectures that push data out of the OT network for AI to use, rather than giving AI systems standing access into it. It sits on top of standards the industry already works to:

ISA/IEC 62443, the consensus standards for industrial automation and control system security, which ISA notes apply to process industries including oil and gas¹⁷.
IEC 61511, functional safety for safety instrumented systems in the process industry¹⁸. Ask your functional safety assessor whether any AI function that touches a safety instrumented system changes the certified design.
NIST SP 800-82 Rev. 3 (September 2023), the US guide to OT security¹⁹, and the voluntary NIST AI Risk Management Framework (January 2023)²⁰.
For pipelines, API Std 1164 on pipeline control system cybersecurity²¹ and the TSA Security Directive Pipeline-2021-02 series²².

Why the control layer is where the risk sits

Three incidents explain why operators are careful about connecting anything new to OT:

TRITON, 2017. Attackers at a Middle East industrial facility reprogrammed Triconex safety instrumented system controllers. Some entered a failed safe state and shut the process down, which is how the attack was found^{23, 24}.
Colonial Pipeline, May 2021. A ransomware attack led Colonial to shut down its pipeline system as a precaution. It restarted six days later²⁵.
A drilling unit in the Gulf of Mexico. In 2015 the US Coast Guard Commandant described malware, brought aboard on workers' personal devices, that disabled the signals to a floating drilling unit's dynamic positioning thrusters. The unit drifted off the well site²⁶.

None of these involved AI. All of them involved the boundary between the systems that run the asset and everything else. Every AI system that reads OT data moves that boundary, which is why the decision about where it sits deserves an owner and a record.

Will AI replace oil and gas jobs?

The primary data says no, not the field jobs, and not soon. The US Bureau of Labor Statistics projects employment of oil and gas workers to grow 2% from 2025 to 2035, slower than average, and says robotics, automated drilling and remote monitoring are "expected to dampen some demand"²⁷. That is slower growth, not replacement.

The International Energy Agency is more direct. In World Energy Employment 2025 it says current AI use cases "do not significantly reduce demand for applied technical workers in construction, operations, and maintenance," because those roles are mostly manual work that AI is not well suited to replace²⁸. The same report counts 12.4 million people in oil and gas fuel production and distribution in 2024, up 210,000 on the year, and finds that across the energy sector, advanced economies have 2.4 workers within ten years of retirement for every worker under 25.

The pressure in the data is the opposite of replacement: a retiring workforce and too few people to replace it. Where AI changes the work is in what one engineer can watch. That makes the record of who decided what more important, not less.

How big is the market?

Estimates vary with what each firm counts. Research and Markets puts AI in oil and gas at $3.54 billion in 2025, rising to $6.4 billion by 2030²⁹. Global Market Insights, counting AI and machine learning together, puts it at $5.0 billion in 2025 and $18.3 billion by 2035³⁰. Read any single figure as one firm's definition, not a market fact.

For the companies selling into it

If you make the automation, controls or software on these assets, the questions on this page are the ones your buyers are working through: where the model sits, who signs off, and what the record shows. ModalPoint helps OT and automation vendors take that conversation to market: positioning, the rooms that matter at the energy conferences, and the visibility that gets you onto the shortlist. The interactive rigs on this site were built for exactly those rooms. See what we do, or start with how operators buy.

For operators

The three rules above are the starting point for AI decision governance: deciding who owns each AI-assisted decision, what gets logged, and how you prove it later. It rests on Digital Information Governance® (DIG), our framework for governing the decision rather than the vendor. If you are working out where AI belongs on your assets, talk to us.

Sources

Book a call

Talk to ModalPoint

A 30-minute call to see if ModalPoint is the right firm and whether the timing makes sense. No obligation either way.

Book a 30-Min Call