Where AI and Machine Learning Are Actually Deployed in Energy

Four labels sort the examples below. In production means a source describes the system as routine operational use. A pilot is a trial, field test, or prototype. Announced covers partnerships, contracts, and launches that come with no evidence of use. Research covers proposals, benchmarks, and testbeds.

Numbers are a separate problem. Savings and accuracy numbers usually come from the operator or the vendor that sells the system, often with no method and no baseline to compare against. Each number below is attributed to whoever reported it. Statements that something was not found are dated, because the record changes quickly.

Forecasting is where the public record shows ML in use the longest. ISO New England, the independent system operator (ISO) for the New England grid, said in an October 2025 industry workshop presentation that it has used a neural-network load forecast, covering up to seven days ahead, for more than 20 years. A load forecast predicts electricity demand. A neural network is a type of ML model built from layered mathematical units that learn from examples. The same presentation called load forecasting the most successful application of AI and ML in grid operations.

The North American Electric Reliability Corporation (NERC) reported similar practice at two other operators in a November 2024 white paper. NERC says the Midcontinent Independent System Operator (MISO) uses neural networks for a seven-day mid-term load forecast and linear regression for a five-minute short-term forecast, and that both forecasts have been active since 2013. For PJM Interconnection, a regional grid operator in the eastern United States, NERC says operators see a blend of ML models and pattern-matching methods, weighted in advance by engineers and forecasting experts, and operators can refine the result using their own experience. A figure in the paper lists four neural-network models, two pattern-matching methods, and two blended forecasts.

MISO also described a second use in a March 2025 board committee briefing. An AI/ML model rates how uncertain the next day's operating conditions are, using inputs such as load, wind, and solar generation. MISO said the model successfully predicted high uncertainty during winter events on January 19 to 22, 2025. The briefing gave no accuracy figures or baseline, so it is a report of use, not a measured result.

The Electric Reliability Council of Texas (ERCOT) described two ML forecasts in an August 2025 review: one for the combined charge held in grid batteries and one for large flexible load. ERCOT says both run live, update every hour, and retrain daily. Large flexible load generally means large electricity customers that can adjust how much power they use, though the review does not define the term. For one typical week, the report shows an average error of about 6 percent for battery charge and about 7 percent for large flexible load, measured as mean absolute percentage error, rising to about 16 percent for large flexible load in the evening hours ending 5, 6, and 7 p.m. The report does not say which forecast horizon the table covers. ERCOT describes a separate ML tool for detecting anomalies in market data as still in development.

Forecasting reaches beyond demand. In Great Britain, the system operator, now called the National Energy System Operator (NESO), reported in July 2024 on a solar forecasting project with Open Climate Fix. The project built a deep-learning model that combines satellite images with weather predictions to forecast solar output up to two hours ahead. NESO said the best national forecast from the project had a mean absolute error of 233 megawatts, compared with 650 megawatts for its existing national solar forecast, and that a first fully operational service reached its control room in December 2022. The page does not say whether the service is in routine use today.

Utilities and oil and gas operators also use ML to watch equipment and flag trouble early. Pacific Gas and Electric (PG&E) said in May 2026 that its new Continuous Monitoring Center draws on data from more than 5.5 million sensors across its grid, with ML models scanning that data continuously. When something looks wrong, a trained team reviews it and decides whether to send a crew. PG&E also reports results, such as potential ignitions intercepted in 2025, but it attributes them to its broader monitoring programs rather than to the new center alone.

Entergy says it applies AI to smart-meter data to predict when distribution transformers are likely to fail. In a 2024 post the company reported that, in one year, the effort prevented 536 unplanned outages. Entergy gave no method or baseline for the figure, so it is best read as a company claim.

In oil and gas, Equinor reported in January 2026 that AI monitors more than 700 rotating machines, using 24,000 sensors, to predict failures. The company attributes $120 million of value to this work since 2020 but does not explain how it calculated that figure. A paper describing a Murphy Oil project, summarized in the Journal of Petroleum Technology in January 2025, shows more of the effort involved. During a 24-month project on two deepwater Gulf of Mexico platforms, Murphy deployed 46 predictive models covering turbines, gas compressors, pumps, and a glycol system. The summary says the scope was to explore and experiment with predictive maintenance, so it reads as a trial rather than routine use. The first model went live six months after the project began, and the rest followed in stages. False alarms rose as models were added, then fell after alerts were reviewed and models retrained. The summary quantifies no savings.

A 2024 summary report from the U.S. Department of Energy (DOE) on AI in critical energy infrastructure lists ten categories of AI applications. It describes only one as already deployed: predictive maintenance for wind turbines and for oil and gas compressors and pumps. It describes operational awareness as something AI is helping system operators do, and it describes the other eight as things AI can or could do. The report summarizes an interim assessment and is not a comprehensive evaluation, and it is now more than two years old.

Methane detection is a regulated use. In September 2025 the U.S. Environmental Protection Agency (EPA) approved a methane monitoring system from the oilfield services company SLB as an alternative method for periodic screening under its oil and gas methane regulations. EPA's memo describes the system's event detection as using a proprietary AI-based algorithm and its quantification as combining ML with a physics-based plume model. EPA's validation covered how small a leak the system can detect, and the memo gives no figures for quantification error. Approval shows the method may be used, not how many sites use it.

San Diego Gas & Electric (SDG&E) described a production use of ML in a March 2026 filing to the California Public Utilities Commission. The filing is retrospective, covering 2019 to 2023. It says ML models check the findings of drone inspections of overhead distribution poles and structures. Its narrative counts more than 60 such models by the end of 2022 and more than 90 in 2023, when every inspection assessment was reviewed both by a supervisor and by the models. A metrics section elsewhere in the filing says 94 machine learning capabilities were developed from 2019 to 2022. The models act as a second reviewer, not a replacement for the inspector. SDG&E also used model predictions from prior-year images to help choose which poles to inspect. The filing cites accuracy of upwards of about 87 percent for automated damage detection in 2023 but does not say how accuracy was measured.

ML also ranks assets by risk. PG&E publishes documentation for ML models that estimate the relative likelihood that equipment such as capacitor banks, fuses, and protective devices will fail or start a fire, and says the results feed a model it uses to prioritize wildfire-risk work. The same documentation rates the capacitor-bank model's performance as very poor and says a physics-based model will be investigated if the ML model keeps performing poorly. A 2012 paper by researchers working with Consolidated Edison described an ML system that had ranked the utility's feeders, its main distribution lines, by failure risk since 2008. By the authors' own evaluation, about 60 percent of failures fell among the top-ranked 15 percent of feeders.

On pipelines, Enbridge says it has used ML models to bring predictive analysis into its more complex defect assessment, which it says prioritizes defects for detailed evaluation. The 2026 customer handbook that says this reports no metrics. Public evidence for ML-based pipeline risk scoring is thinner. As of October 2026 the main source we found is a research project funded by the Pipeline and Hazardous Materials Safety Administration (PHMSA) that ended in September 2025. Its final report, posted in a revised version dated January 2026, comes from researchers at Rutgers and Marquette and describes research models tested on inspection data from two pipeline operators, not an operator deployment.

Inspection with drones and robots is at an earlier stage in oil and gas. Chevron describes a six-month pilot of AI-equipped drones for facility inspection, followed by a second system in Colorado in 2025, and says it is evaluating more. Petrobras announced in 2023 that it was buying six inspection robots for testing and listed automated AI checking of inspection images as a later step. Neither reports measured results.

Equinor also reports that an AI tool gave it a tenfold increase in seismic interpretation capacity, with 2 million square kilometers interpreted using it in 2025. The company does not say which interpretation tasks the tool performs. Separately, Equinor's executive vice president for Technology, Digital and Innovation says the company primarily uses traditional ML on its operational data. Shell lists AI-assisted sorting of seismic data among its uses but gives no metrics for that use.

On wells, the Abu Dhabi National Oil Company (ADNOC) said in July 2024 that RoboWell, a cloud-based AI well-control system from the company AIQ, had been deployed at an onshore field in 2023 and across an initial 10 wells at an offshore field, with deployment scheduled for more than 300 wells across offshore and onshore operations in 2024. The quantified benefits, up to 30 percent optimization in gas-lift consumption and up to 5 percent increase in operating efficiency, were stated by AIQ's chief executive, not demonstrated in the release. Gas lift is a method of bringing oil to the surface by injecting gas into the well.

Not every advanced automation system is described as AI by the company that built it. In June 2025 Chevron and Halliburton announced a closed-loop hydraulic fracturing workflow in Colorado, in which automated stage execution adjusts to real-time subsurface feedback without relying on human intervention. Their announcement describes automation and an algorithm that carries Chevron's engineering knowledge. It does not use the term AI, and it reports no results.

Much of the control equipment that runs energy systems is long-established automation that follows defined rules or physical models. Examples are protective relays, which detect faults and trip breakers, automatic generation control, which adjusts power plant output in response to changes in load, frequency, and scheduled flows between areas, and state estimation, which calculates the most likely present condition of the grid from imperfect measurements. NERC's white paper cites the state estimator as an existing tool of the energy management system, the software control rooms use to monitor and control the grid, and it reports that interview participants pointed to tools advertised as AI or ML that in fact used other techniques, such as simpler statistics.

Advanced process control, usually model predictive control, which uses a model of the process to plan control adjustments ahead of time, is another case that vendors treat as distinct from AI. Honeywell's 2021 white paper describes it as multivariable model-based control and presents AI and ML as newer methods that complement it, and ARC Advisory Group wrote in April 2026 that the two paths are converging. The European Commission's 2025 guidelines on how to define an AI system under the European Union's AI Act make a related point. They say the key test is whether a system infers how to generate outputs from its inputs, that this can be done with machine learning or with logic- and knowledge-based methods, that some simple ML-based systems still fall outside the definition, and that the system need not keep adapting after deployment. The guidelines are non-binding. This is why the examples in this article stay close to what each source says the system does.

Generative AI is newer, and the public record shows mostly early use, usually with a person checking the output. Retrieval is one of the most common uses in the public examples we found: searching large collections of manuals, procedures, and records and returning answers. Some of these tools, such as the one Tapestry describes for PJM below, use retrieval-augmented generation, in which the system first looks up relevant documents and then asks a language model to write an answer from them.

PG&E and the AI company Atomic Canyon announced in November 2024 that generative AI would be deployed at the Diablo Canyon nuclear plant, starting in early 2025, to search technical documents. In December 2025 Oak Ridge National Laboratory quoted a Diablo Canyon vice president as saying the plant had already seen increased productivity in certain areas, and quoted the plant's AI project director naming procedures, training, and assessments as next steps. As of October 2026 we found no measured plant results.

Chevron says its in-house generative-AI platform, ApEX, launched in August 2024 for exploration work. Several search agents, each focused on a topic such as geospatial data, search more than a million files and return insights to exploration teams. Chevron says the platform saves time and cost but gives no figures.

Drafting is earlier still. A December 2024 presentation from Idaho National Laboratory described a pilot at a reference nuclear plant, under DOE's Light Water Reactor Sustainability program, in which a chatbot trained on plant documents such as work orders (records of maintenance tasks), procedures, and past evaluations drafts equipment evaluation reports that engineers then validate. The presentation lists time saved and quality as open research questions. Separately, the Electric Power Research Institute (EPRI) listed field tests of voice assistant and AI tools at Southern Company, FirstEnergy, and the New York Power Authority for work order creation, data retrieval, and safety communication. EPRI said the tests reduced administrative burden and improved data accuracy, without giving figures. It does not say whether the tools use generative AI.

Grid operators are also testing AI on interconnection, the process of studying and approving new power plants that want to connect to the grid. PJM said in April 2026 that HyperQ, an AI-enabled tool developed by Google's Tapestry, is already helping with application reviews by identifying sections of documents for further review, and that it will evaluate the tool's effect during the current application cycle. Tapestry, the vendor, reports that PJM engineers make the final determinations. Other efforts are less clear. The Southwest Power Pool (SPP) announced a development partnership with Hitachi in June 2025 with a goal of cutting interconnection study times by 80 percent, and as of October 2026 we found no report of production use. In 2025 MISO benchmarked SUGAR (Suite of Unified Grid Analyses with Renewables), a study-automation tool from Pearl Street Technologies, and no source we read describes it as AI.

The California Independent System Operator (CAISO) described a prototype in July 2025 of an AI assistant from the software company Open Access Technology International (OATI), intended to help verify planned-maintenance requests. CAISO said operators will keep final authority and the suggestions will remain advisory. OATI has since said the assistant has run in CAISO's production outage management system since June 24, 2026. As of October 2026 we found no statement from CAISO or an independent source that confirms this.

Public evidence for control-room tools is mostly at the pilot stage, though Honeywell has announced a commercial launch of one. TotalEnergies reported in November 2025 an ongoing pilot of Honeywell's AI-assisted Experion Operations Assistant at its Port Arthur refinery in Texas, where an initial pilot ran at the delayed coking unit, which processes the heaviest refinery residue. In preliminary results the tool forecast five potential events, on average 12 minutes before an alarm incident. The release does not say who measured this. It says the forecasts help minimize downtime and reduce emissions from flaring but gives no figures for either. Honeywell announced a commercial launch of the product in March 2026, citing pilots with Chevron and TotalEnergies and an average lead time of 5 to 10 minutes across its pilots.

On the electric side the evidence is earlier. EPRI showed an LLM-based tool for asking questions about alarm data at a March 2026 webinar, and its slides describe it as early stage. Researchers at the National Laboratory of the Rockies, formerly the National Renewable Energy Laboratory, have shown similar tools as research prototypes and designs for future control rooms. Neither source states that a utility uses such a tool in operations.

AI agents that plan and act across systems are the newest category, and the public record is mostly announcements. In September 2026 Amazon Web Services (AWS) announced a program in which Duke Energy engineers are now working with AI agents on data preparation and analysis runs for interconnection studies, with engineers making the decisions. AWS says data preparation went from two weeks to hours, a company claim, and it does not say whether the work is a pilot or in production. EPRI's own slides label its protection-agent work as current research and development experiments. POWER Magazine reports that nearly every grid-related award it describes under the Department of Energy's Genesis Mission research program is a Phase I, or first-stage, effort.

In oil and gas, an AWS case study reports that the pipeline company TC Energy built an agent called PIPER that checks pipeline material test reports against weld maps and pressure-test documents. The case study says it has more than 100 active users, cut document review time by 85 percent, and reached 95 to 100 percent accuracy in early testing on midsize documents of up to 120 pages, with engineers doing the final review. These are figures published by AWS. The case study also says PIPER and a second agent together are expected to deliver more than $18.7 million in avoided maintenance capital costs over a five-year planning horizon. That is an expectation, not a result. The TC Energy staff quoted on the page do not repeat these figures, and we found no TC Energy publication that does.

Autonomy in control is rarer still. The 2024 DOE summary said its assessment, based on input from energy sector entities and experts, indicated that autonomous AI is not likely to be implemented for active control in critical energy systems in the near term, given the potential consequences of errors. It said AI-assisted decision support with a human in the loop, meaning a person reviews the AI's output, appeared to be a promising alternative. NERC's 2024 white paper describes three modes for AI in real-time operations: active control, co-pilot, and decision-making support. It says the operator should have the final input on AI-generated actions, and it notes that active control can run with or without human approval. A NERC subcommittee took up the question again in March 2026, when it asked the Reliability and Security Technical Committee for 30-day comments on a white paper on an Electric AI/ML maturity model. The meeting presentation says the paper's automation scale treats decision support, human-in-the-loop use, and bounded autonomous operation as separate patterns. The paper itself is not in the meeting package, which predates the meeting, and we did not find minutes showing what the committee decided.

A few narrow, single-purpose systems do act on equipment. ADNOC's RoboWell and the Chevron and Halliburton fracturing workflow, described above, are examples, though the releases do not describe their limits or any human override. In June 2026 Honeywell announced Experion Cognition, which uses agents to detect and help resolve abnormal process conditions, and described a live proof of concept at a Borouge petrochemical site in Abu Dhabi. The release gives no results from the proof of concept. Its language is stronger, saying the product brings autonomous control rooms into production, but the deployment it describes is a proof of concept. Honeywell said the product would be commercially available in the third quarter of 2026. As of October 11, 2026 we found no confirmation that it has been released.

The clearest accounts of what slows deployment come from operators themselves. PJM published a 2024 paper reporting ML projects that were stopped, kept out of production, or struggling. A model for predicting net interchange, the net power flowing between neighboring systems, performed well on some interfaces, but the team had no dedicated staff to evaluate and maintain it, so the project ended. A model that read outage tickets needed too much data preparation, raised the risk of bias, and covered a process PJM decided should stay with human judgment. Two anomaly-detection efforts produced results that gave no guidance on what to do. The paper does not name any ML model as in production, though it says PJM has used forms of AI for some time.

The Great Britain system operator's own close-down report, filed in July 2024 before it became NESO, shows a similar pattern. A proof of concept that used AI to triage outage requests was closed out under the Network Innovation Allowance, a limited innovation fund within the network price control overseen by Ofgem, Great Britain's energy regulator. It showed large discrepancies and cost estimates that were not credible, and the operator parked it. ERCOT lists continuous model tuning, data governance, and security among the challenges to wider adoption.

Rules also shape adoption. NERC's Critical Infrastructure Protection (CIP) standards apply mandatory cybersecurity and physical security requirements to qualifying parts of the bulk electric system, the interconnected high-voltage grid, scaled by impact. As of October 2026, our keyword scan of the latest listed versions of the 14 CIP standards, CIP-002 through CIP-015, found no mention of AI or ML in the standard text. Thirteen of those 14 latest versions are not yet enforceable. Only CIP-012-2 is. In June 2025 the Federal Energy Regulatory Commission (FERC) approved CIP-015-1, a standard for monitoring traffic inside certain networks, in Order No. 907. It does not prescribe a particular technology.

Guidance is mostly voluntary. In the United Kingdom, Ofgem published non-binding guidance on ethical AI use in the energy sector in May 2025 and a second edition in May 2026. In Arizona, Commissioner Lea Márquez Peterson opened a formal inquiry into utility AI use in March 2026. NewGen Strategies reported that, as of October 2026, no state utility commission had issued formal guidance on utility operational AI, and it found only Arizona and New York with formal proceedings.

Assurance practice is still forming. In 2026 DOE's Office of Cybersecurity, Energy Security, and Emergency Response, working with Lawrence Livermore National Laboratory, announced the Mjölnir testbed in April for adversarial testing of energy AI models and the Stormbreaker testbed in July for evaluating LLMs and agents in power-system and operational technology (OT) environments. OT is the equipment and control systems that run physical processes. EPRI announced SAFERai.power in August 2026, an initiative with more than 25 founding members to develop a risk framework for AI in the power sector. The public pages for these efforts describe capabilities and goals, and as of October 2026 they report no evaluation results. EPRI's roadmap puts its first published version, v1.0, at months 13 to 18, with the full open-source toolkit targeted for 2028.

Several uses that vendors and conference programs discuss have no public operator evidence behind them. As of October 11, 2026, we found no operator or regulator document describing ML price forecasting in routine use, ML leak detection on pipeline networks, AI-drafted work orders or inspection reports, or AI correction of grid models or contingency screening, which means checking how the grid would behave if a line or generator failed. Related uses exist, and they differ from the ones listed. The facility-level methane monitoring described above is not pipeline leak detection. The Idaho National Laboratory pilot, which drafts equipment evaluation reports, and the EPRI voice-tool field tests are the nearest examples we found for drafting. ISO New England's slides describe an ML tool that estimates transmission-line outage probability from weather forecasts and feeds a cascading-failure analysis, which is related to contingency analysis but is not described as screening. We searched government, laboratory, regulator, operator, and conference sources, and some evidence may sit behind paywalls or in unpublished work.

Two more limits apply. Savings and accuracy figures come mostly from the operators or vendors that stand to benefit, and many come with no method or baseline. And a source that says a system is available, or that a pilot is complete, does not say how many sites use it. For both reasons, treat any single sector-wide adoption figure with suspicion.