Energy: AI reaches the control room as the EU delays its deadline

Energy is the industry where AI is arriving inside the control systems operators already own. Since Era Haus last looked, Emerson announced AI agents for the software that runs many power plants. Iberdrola says it has close to 500 agents across its networks, power stations and customer service. The European Union, meanwhile, moved its compliance deadline for AI in critical infrastructure to December 2027. The tools moved closer to the work, the evidence that they pay stayed where it was, and the regulatory clock gained sixteen months.

AI arrives inside the software you already run

Our last look, Energy: the power grid fell short, and the backlash began, covered the electricity AI consumes; this one covers AI inside the business. The clearest move came from Emerson, a US control-systems maker, in its own newsroom on 27 July 2026. Its Ovation platform, which runs many power and water plants, will carry five AI agents, programs that each handle one job on their own, due later this year, with 28 more planned. One explains why a control sequence stalled. One sorts and summarizes alarms during an alarm flood, when so many fire at once that no operator can read them all. The rest cover predictive maintenance, tracing faults to their cause and finding badly tuned controls.

KROHNE, an instrumentation maker, launched PipePatrol NEO at a Berlin pipeline conference in May 2026, and the trade outlet Pipeline Technology Journal covered it on 28 July 2026. It runs a live simulation of the pipeline and compares real pressure and flow against it, flagging leaks and theft with fewer false alarms. It fits onto existing instruments and SCADA, the software a control room uses to watch and steer equipment remotely.

For oil and gas, the unit being sold shrank. Gain.Energy, a US software company, announced Upstrima in a press release on 18 August 2026: a subscription web application where an engineer picks one agent for one task, such as end-of-well reporting or stuck-pipe prevention, keeping the drill pipe from jamming in the well. It reads more than 50 file formats, including scanned and handwritten documents, where most small operators' well records still sit. Its claim of 50 to 70 percent faster work names no customer.

Now it arrives as a feature of what you already run, and the decision becomes an upgrade, a license or a subscription.

The biggest claim came with no numbers

Iberdrola, Spain's largest electricity company, said in a press release on 17 August 2026 that it has close to 500 AI agents deployed or in development: on the grid, anticipating faults and restoring supply after outages; in customer service, as automated phone assistants handling questions and complaints; and in corporate functions and IT.

What Iberdrola did not publish matters as much: no savings figure, no vendor, no model, no headcount. It said people keep responsibility for decisions and paired the count with training. The 500 is a count of activity. Whether any of it pays is a question Iberdrola has not answered, and a smaller operator citing it as proof of return is citing a claim Iberdrola never made.

None of these launches names the model underneath, so none can be checked against the best current models, and every efficiency number on offer comes from the seller. No Latin American operator published a verifiable operational deployment in the same weeks.

The EU moved its deadline days before it applied

The European Union published its Digital Omnibus on AI, the package of amendments to the EU AI Act, on 24 July 2026, days before the deadline it was written to move. Obligations for high-risk AI in critical infrastructure were due from 2 August 2026. They now apply from 2 December 2027.

The list to check your systems against comes from a May 2026 analysis by the law firm Baker Botts: grid management, dispatch (deciding which plants run when) and fault detection; pipeline control and leak detection; automated well control; safety monitoring at liquefied natural gas plants. The tools launched since July are the tools the rule covers.

Elsewhere nothing moved. Ofgem, the UK energy regulator, has published nothing since its May 2026 AI guidance. In the US, the policy outlet TechPolicy.Press found in April 2026 that no state utility regulator had yet ruled on whether spending on operational AI can be recovered through customer rates.

What it means for you

If you run Ovation or a standard SCADA, the AI conversation is now with the vendor you already pay, and the cost moves from a project budget to a line item. The risk moves too, from integration, which you can scope, to dependence on one vendor's roadmap.

The first jobs these agents take are alarm triage, reports and procurement analysis: the repetitive reading and sorting that fills a shift. Nothing since July shows an energy operator cutting staff because of AI; the one that published a number paired it with training. Plan for shifts that read less and check more.

Compliance is now a schedule. In the EU, 2 December 2027 is the date and the list above is the checklist; sixteen months is enough to document who decides what and how a person overrides the tool, and too little to start in mid-2027.

What to do about it

First, ask your control-system and instrumentation vendors what agents they plan and whether a named customer will speak to you. Run the first one on alarm triage with a person checking every output, and measure your own before and after; that number will be worth more than any vendor's. Second, map your AI use against the EU's high-risk list this month; in the EU it is law with a date, elsewhere a free self-audit.

On the claims, wait for a named customer with a measured result. What changed since July is the route AI takes into an energy business: through the vendor you already pay, for a job you already do. The proof still has to come from you.