Energy: where the industry stands with AI

Where energy stands with AI

Energy in 2026 uses AI mostly where it has for years: software that reads sensor data and warns of a failing pump, a damaged line or a leaking pipe before it breaks. What is new is AI agents, programs that carry out a task on their own, arriving inside the control, planning and customer systems operators already run. The proof trails the promise. Most large operators have something running; few can show a measured return, and those that can got most of it from maintenance.

Adoption is wide and thin. Bain, the consulting firm, surveyed more than 800 energy and natural-resources executives and published in September 2026: most expect AI to matter a great deal within a decade, and fewer than 20% report measurable results at scale today.

Among US utilities, a survey of 134 innovation leaders by National Grid's venture arm, published September 18, found 78% deploying at least one AI application to handle the growing queue of requests to connect new projects to the grid. In a spring 2026 survey by the accounting firm Grant Thornton, 17% of US energy leaders said they were ready for an independent audit of how they govern AI.

What works today is predictive maintenance, image inspection, wind and solar forecasting and customer email. What is still promise is agents that plan a grid or run a drilling program with little supervision, and any measured saving from them. None of the tools named in this report says which AI model it runs on, so no efficiency claim here can be checked against the best current models.

What changed in the last two months

New York made AI use a regulatory filing. On September 17, 2026, the New York Public Service Commission, which sets the rules and rates for the state's utilities, ordered the state's electric, gas and water utilities to file an inventory of every AI system they use, with the policies that govern it, by mid-November and every six months after (Utility Dive). Arizona's utility regulator opened a similar inquiry in March. Era Haus covered the order in Energy: New York orders utilities to list every AI tool they use. If you sell services to a US utility, expect to be asked what AI your work uses and who checks it.

Agents reached the engineering desk. On September 17, Amazon Web Services launched agents for interconnection studies, the engineering checks a utility runs before a new solar farm, battery or factory can connect, built with Duke Energy. Duke says data preparation fell from about two weeks to hours; no one outside the two companies has measured it (T&D World). Iberdrola, Spain's largest electricity company, said on August 17 it has close to 500 agents live or in development, with no savings figure attached.

One of the largest operators is building its own models. TotalEnergies committed more than €100 million over three years with Mistral, the French AI company, to build models for exploration and for getting more out of existing oil and gas fields, expanding an earlier partnership (Bloomberg, September 15). Nothing has been delivered yet.

Europe's deadline moved. The EU's Digital Omnibus, a package of changes to its digital laws published July 24, 2026, pushed back the AI Act's extra obligations for "high-risk" AI (risk checks, human oversight, record-keeping) in critical infrastructure such as energy supply, from August 2, 2026 to December 2, 2027.

What works in practice

Start from the sensors you already have. Equinor, Norway's state-controlled oil and gas company, says AI brought it $130 million of savings and added value in 2025, and about $120 million of that came from watching more than 700 pumps, compressors and other rotating machines through their sensors (Equinor, January 2026). The dependable return is in plain predictive maintenance: years of instrument data turned into early warnings.

Pick two or three jobs and measure them. Bain found that the companies getting results concentrate on a few areas, such as equipment performance and technical services, and judge each tool by a business number.

Buy together. American Municipal Power, which buys electricity for more than 130 US towns that run their own utilities, offers its members a camera system mounted on ordinary trucks that photographs poles and flags defects, at group pricing (American Public Power Association, April 2026). German municipal utilities buy the same way through Thüga, their shared services network.

Label the machine and keep a human route. Octopus Energy, a UK electricity and gas supplier, let an AI assistant answer about 8,000 routine customer emails a week for three months. Customers rated 76% of its replies satisfactory, against 72% for comparable replies from staff (Octopus, July 2026). Every AI email was labeled, and any customer could ask for a person.

Write it down and check the vendor. NRECA, the US association of electric cooperatives, updated its sample AI policy in July 2026 and tells members to list the tools their staff use. And a multi-year contract is only as good as the company behind it: C3 AI, a US AI software company that has run Shell's maintenance program since 2018, saw revenue fall about 36% in its fiscal year to April 2026 and cut about a quarter of its staff (company filings).

On jobs, none of the large oil companies that cut staff in 2025 and 2026 blamed AI; they cited cost-cutting and mergers.

What named companies are doing

PG&E, the northern California utility, opened a monitoring center in May 2026 that runs machine learning over grid sensors and about 5.5 million smart meters. The center houses monitoring the utility already ran, and PG&E credits that work with intercepting 17 potential wildfire ignitions in 2025 and about $6 million in operating savings (PG&E, T&D World). The safety gain is worth more than the savings.

Shell extended its maintenance program with C3 AI in June 2026. It covers more than 13,000 pieces of equipment, and the new stage adds agents that propose the cause of a fault and the fix. Shell has published no result for it.

In Argentina, YPF's La Plata refinery installed an AI system, built with the consulting firm BCG, that tracks fuel components in real time to blend gasoline and diesel. YPF reports blending time cut by up to half against an 18-to-20-hour lab-sample routine, and the system won a Latin American refining award in Buenos Aires in September. In Chile, the Spanish renewables company ACCIONA Energía says an AI-based control system on southern transmission lines raised their capacity by 60%, letting more wind power through (IndustriAmbiente, August 2026). Both figures are the companies' own.

The tools, by size of operator

Corporate. Large operators get AI inside platforms they already license, as Shell does with C3 AI. Coming: Emerson announced in July 2026 AI agents for Ovation, the control system in many power and water plants, to sort alarm floods and explain stalled control sequences; the first are due by the end of 2026. Enverus ONE, launched in April 2026 by Enverus, a US oil-and-gas data company, lets US oil and gas teams query Enverus's well and lease data in plain language and pull clauses out of lease documents. In Argentina, YPF renewed its contract with Corva, a US drilling-data software company, in September 2026 for a center where engineers watch live data from its wells.

Small business. The tools arrive through the vendor or the buying group. US municipal utilities get the truck-mounted pole inspection, from a company called Noteworthy AI, through American Municipal Power. In Germany, about 100 municipal utilities can buy a private AI assistant from Deutsche Telekom under a framework agreement Thüga signed (Telekom, April 2026). Residential solar installers meet AI inside the design software they already pay for: Aurora Solar, used mainly in the US, drafts designs and customer proposals and added heat pumps and batteries in September 2026. OpenSolar, strongest in Australia and the UK, added an assistant in January 2026 that builds a design from spoken commands.

Freelance and solo. Not enough verified evidence this version on what independent energy consultants, engineers and one-person installers use; the next update will add it.

What to watch over the next year

First, the New York filings. The inventories are due in mid-November 2026. Watch whether the commission moves from a list to rules on human oversight and cybersecurity, and whether any US regulator decides if utilities can charge customers for their AI spending through rates.

Second, how narrowly Europe defines high-risk. The European Commission's draft guidelines of May 19, 2026 would cover energy AI only when it is a safety component in running electricity, gas or heat supply and the operator is one its national government has formally listed as essential to energy supply; tools that only inform or optimize fall outside. If the final text keeps that line, most AI in a small EU energy firm carries no high-risk duties, before or after December 2, 2027. In Germany the grid regulator, the Bundesnetzagentur, will also enforce the AI Act under a law passed in June 2026, and in the UK the energy regulator Ofgem starts a 12-month AI sandbox, a supervised trial space, in late 2026.

Third, the fine print. Since January 2026, ISO, the body that writes standard US business-insurance forms, has offered insurers optional clauses that exclude harm caused by generative AI, and carriers are adding them at renewal (Insurance Journal, August 2026). In EU wholesale energy markets, any firm that trades by algorithm, AI agent included, must notify its national regulator and ACER, the EU agency of energy regulators, under the revised market-abuse rules; ACER's second-quarter 2026 report shows more cases of suspected market abuse involving trading algorithms.