EU AI Act for energy utilities: what applies after the Omnibus

Which dates apply now

Since the Omnibus, four dates matter for energy utilities: February 2025 for prohibitions and AI literacy, August 2026 for transparency, December 2027 for high-risk AI under Annex III and August 2028 for AI in regulated products under Annex I. Regulation (EU) 2026/1744 was published in the Official Journal on 24 July 2026 and has applied since 27 July. Parliament and Council reached the political agreement on 7 May 2026.

Anyone who built a 24-month plan towards August 2026 in 2025 gains 16 months for the high-risk parts. For the chatbot notice and training they gain nothing. Those already apply.

How dates, risk classes and duties fit together is set out on one page in our AI regulatory matrix for energy utilities.

When energy AI is high-risk

Under Annex III point 2, AI at an energy utility is high-risk only if it works as a safety component in operating electricity, gas or heat supply. The European Commission's draft guidelines of 19 May 2026 attach two conditions to this, and both must be met. The draft is not final. The Commission has announced the final version for the end of 2026.

• The AI directly protects the physical integrity of the installation. Systems for forecasting, efficiency or information do not qualify under the draft.
• The operator has been designated by a Member State as a critical entity under the CER Directive. In Germany this is governed by the KRITIS-Dachgesetz, in force since 17 March 2026; the ordinance with the thresholds, which decides KRITIS status and with it high-risk AI at grid and heat operators, was still in progress in September 2026.

What remains is a short list. Most AI use at German municipal utilities, from sales forecasting to clearing cases in market communication, sits outside it. The tricky part is AI that switches on its own without an independent protection layer behind it. The examples and grey zones are covered in which AI at a utility is really high-risk. The obligations these systems face from December 2027 are set out in high-risk AI in energy infrastructure.

Two more Annex III entries that can affect utilities

Besides critical infrastructure, Annex III lists seven other areas. Two of them touch utilities. Point 4 covers AI in recruitment and workforce management, such as a tool that pre-sorts applications. Point 5(b) covers AI that assesses the creditworthiness of natural persons; a utility using such scoring before signing a household customer is the deployer of a high-risk system. Algorithmic pricing, by contrast, is not in Annex III, contrary to what the first version of this article said.

Provider or deployer: the role decides

Most obligations attach to the role, not the system. A provider develops an AI system, or has it developed, and places it on the market under its own name. A deployer uses it under its own authority in a professional context. A municipal utility that buys forecasting software or a chatbot is almost always a deployer.

The role can flip. Under Article 25 a deployer becomes a provider if it markets a high-risk system under its own name, substantially modifies it, or changes the purpose of a system so that it becomes high-risk. In-house builds for the grid control centre are exactly this case.

What deployers of a high-risk system have to do

• Use the system in line with the provider's instructions for use (Article 26(1)).
• Assign human oversight to people with the competence and authority for it.
• Check input data, where the deployer controls it, for relevance and representativeness.
• Monitor operation, inform provider and authority of risks, and keep automatically generated logs for at least six months.
• Inform workers' representatives and affected staff before use in the workplace (Article 26(7)).

In Germany the last point meets the Works Constitution Act (BetrVG). Since 2021 it mentions AI explicitly, in sections 80(3), 90(1) no. 3 and 95(2a); co-determination on technical monitoring under section 87(1) no. 6 applies anyway. For AI co-determination, municipal in-house enterprises fall under state staff representation law instead.

Two duties work differently for grid AI. The fundamental rights impact assessment under Article 27 does not apply to systems under Annex III point 2, but it does apply to credit scoring under point 5(b). And for high-risk systems in critical infrastructure, Article 49(5) provides for registration at national level rather than in the public part of the EU database.

What already applies: Articles 4, 5 and 50

Three sets of duties apply regardless of any high-risk classification. They hit almost every utility using AI, including one that only uses a writing assistant in the office.

Article 4: AI literacy

Since 2 February 2025 providers and deployers have had to look after the AI literacy of their staff. The Omnibus softened the duty. Instead of ensuring a sufficient level of literacy, Article 4 now requires providers and deployers to support the development of AI literacy among their staff; a fixed level is no longer in the text.

Less strict is not optional. If you ever have to explain why someone trusted a model blindly, a documented training record helps.

Article 5: prohibitions

Also prohibited since February 2025 are, among others, social scoring and emotion recognition in the workplace, except for medical or safety reasons. For utilities the second one matters: software that analyses the mood of your own customer service agents is not allowed. Recognising the mood of callers is not prohibited, but it falls under the information duty in Article 50(3).

Article 50: transparency since 2 August 2026

Since 2 August 2026 people must be able to tell that they are talking to AI. For a customer service chatbot that means the notice belongs in the dialogue, not in the legal notice. Legally the duty sits with the provider of the system. The trouble lands with the utility whose logo is in the chat window anyway.

Anyone publishing AI-generated text on matters of public interest has to label it unless a person has reviewed it and taken editorial responsibility. Generative systems that were on the market before 2 August 2026 have until 2 December 2026 for machine-readable marking of their outputs.

And the large language models? Their obligations have applied since August 2025, but to the model providers. A utility using such a model through an interface does not become a model provider.

Supervision in Germany: KI-MIG and BNetzA

Since 29 July 2026 the Federal Network Agency (BNetzA) has been market surveillance authority, single point of contact and complaints body for the AI Act under the KI-MIG, Germany's implementing act. It is directly responsible for AI in critical infrastructure, human resources and education. In fully harmonised product areas the sector supervisors stay in charge, such as BaFin in financial services.

Two of the agency's offers are useful in practice: the KI-Service-Desk with an online tool for a first assessment of the risk class, and an AI regulatory sandbox for supervised testing, aimed mainly at small and medium-sized companies. The tool does not replace a classification. It does show how the authority asks its questions.

The KI-MIG adds no new substantive obligations. What the authority may request and order is covered in KI-MIG: supervision of your AI now has an address.

Fines and security law

Article 99 sets the frame: up to 35 million euros or 7 percent of worldwide annual turnover for prohibited practices, up to 15 million euros or 3 percent for breaches of most other obligations, up to 7.5 million euros or 1 percent for incorrect information to authorities. For SMEs the lower amount applies in each case.

The AI Act does not replace security law. Since 6 December 2025 the new BSI Act from Germany's NIS2 implementation has applied. It has no AI section of its own; AI systems are covered by general risk management there. That is our reading of the law, it is not spelled out.

What utilities should do now

The order follows from the dates: first what already applies, then the short list of high-risk candidates.

From now until December 2027

1. Build an AI inventory

   List every system with an AI function, including bought-in ones: the forecasting module in energy data management, the chatbot, the writing assistant, functions in the control system. Without a list there is no classification.

2. Record the role per system

   Provider or deployer, with reasons. Check in-house builds and heavily customised systems separately.

3. Check chatbot and published texts

   Applies since August.

4. Training by role

   Grid control, customer service and HR do not need the same training. Keep a record of who got which.

5. Clarify critical-entity status, mark candidates

   Whether Annex III point 2 can apply at all depends, under the draft guidelines, on status as a critical entity. Clarify that first. Then mark the systems that directly protect or switch on their own, and write down the reasoning even when the answer is "not high-risk". When the final guidelines appear, go through the list a second time.

6. Prepare high-risk candidates

   Request documentation from the provider, name oversight staff, set up logging, involve workers' representatives. By 2 December 2027.

Frequently asked questions