Energy: New York orders utilities to list every AI tool they use

Energy utilities in New York now have to write down and file every AI system they use. On 17 September 2026, the state's utility regulator gave eleven electric, gas and water companies 60 days to list each one, with the rules that govern it. In the same weeks, Duke Energy put AI agents on grid-connection studies and TotalEnergies committed more than €100 million to exploration models of its own. AI is reaching the engineering desk of the energy business, and regulators now want to see the list.

New York wants the list

Since our last look, Energy: AI reaches the control room as the EU delays its deadline, the biggest change came from a regulator. The New York Public Service Commission, which sets the rules and rates for the state's utilities, voted unanimously on 17 September to open a formal review of how they use AI. The order names eleven companies, Con Edison and National Grid's New York businesses among them. Each must file, by mid-November, a written inventory of "all use cases of AI systems in their operations", with the policies and controls that govern them, and update it every six months. Utility Dive and RTO Insider covered the order in the days after.

The order lists what worries the commission: wrong answers an AI tool invents, bias, privacy, misconfiguration, cyberattacks, and systems that break when conditions change. It also says the extent of AI in New York utility operations "remains unclear". The regulator does not know what is running, and it suspects some utilities do not fully know either. Reports on the order cite uses already public, such as National Grid estimating spare grid capacity for new connections and the New York Power Authority reading drone footage of trees near power lines.

No other US state regulator took a similar step in the window.

Agents reach the engineering desk

On 18 September, Amazon Web Services launched AI agents built with Duke Energy, one of the largest US utilities, for interconnection studies: the engineering checks a utility runs before a new solar farm, battery or factory can connect to the grid. The agents work inside the utility's existing simulation software and follow its standards. AWS says Duke cut the data preparation from about two weeks of manual work to hours. That is the vendor's number; Duke has not published one.

The same day, a survey by National Grid's venture arm found that 78% of the 134 US utility innovation leaders it asked are deploying or putting into use at least one AI application to handle the rising number of connection requests. It names no tool and no result per utility.

For small generators, the useful launch came from Google. Its DeepMind lab released WeatherNext 3 on 3 September: hourly global weather forecasts, including wind at turbine height and sunlight at ground level, delivered as data through Google's cloud. It is aimed at grid operators, energy traders and renewable developers. No energy customer and no accuracy comparison against the paid forecast services it competes with has been published.

Operators build and buy

TotalEnergies announced on 15 September a three-year program worth more than €100 million with Mistral, the French AI company. The two will build models for oil and gas exploration and for extending the life of existing fields, from a joint lab; Bloomberg reported it the same day. Nothing has been delivered yet.

YPF, Argentina's state-controlled oil company, went the other way and bought. It renewed its contract with Corva, a US drilling-software company, for its real-time center, a room where engineers watch live data from its wells. Those wells are in Vaca Muerta, the large shale oil and gas field in Patagonia, and the roadmap adds alerts, assistants and agents, according to several regional papers on 1 September. Later that month, YPF's La Plata refinery won a digital-technology award at a Latin American refining conference in Buenos Aires for an AI system that tracks fuel components in real time to blend gasoline and diesel. YPF published no yield or margin figures.

At the other end of the scale, the IT arm of Thüga, a network of German municipal utilities, presented a shared AI agent platform in September, hosted in German data centers. It answers staff questions from internal documents and checks supplier offers against pre-agreed supplier contracts, according to the German trade outlet stadt+werk.

What it means for you

In New York, an AI inventory is now a filing. If you sell software, engineering or field services to a New York utility, expect questions about what AI your work uses, who checks its output and what happens when it fails. Outside New York, treat the order as a template. State regulators often borrow from each other, and the EU's rules for high-risk AI in critical infrastructure will ask for similar records from December 2027.

Interconnection study preparation is billable work for many small engineering firms. If the AWS claim holds even partly, those hours shrink, and so does what a client will pay for them.

The largest operators are building their own tools. A small operator will get AI through the contractors and software it already pays for, so the questions go to those suppliers.

What to do about it

Write your company's AI inventory this month, whatever your country: each tool, what it drafts or decides, who checks it, which vendor supplies it. One page covers most small businesses. It answers the question New York now asks and starts the paperwork the EU will want.

If you own solar or wind assets, run WeatherNext 3 next to your current forecast provider on your own sites for a month before changing anything. On AI agents for engineering studies, watch and wait for a utility, not its vendor, to publish the time it saved.