EU AI Act: High-Risk AI in Energy Infrastructure

What the Digital Omnibus actually moved

Start with the plain version. 2 August 2026 is off the table. For high-risk AI under Annex III the date is now 2 December 2027. That sounds like breathing room. It is not.

The European Commission proposed the Digital Omnibus on 19 November 2025, Parliament and Council agreed the text, and since 24 July 2026 it has been in the Official Journal as Regulation (EU) 2026/1744, in force since 27 July 2026. What moved is the day the duties bite. The duties themselves are unchanged. Confuse the two and you plan the deadline wrong.

It was close, though. On 28 April 2026 the second trilogue broke down after about twelve hours over a single question: should AI in regulated products such as medical devices and machinery fall under the AI Act on top of the sectoral conformity assessment? Agreement came on 7 May, and on 13 May the member states confirmed it in the Council. The dispute was settled by giving Annex I its own, later date of 2 August 2028. For grid AI that is a side issue: it falls under Annex III, if at all.

The reason for the delay is unspectacular. The harmonised standards and support tools that make practical implementation possible were not ready in time; the first, EN 18286, only appeared in the Official Journal on 14 October 2026. The deadline follows the standards, not the other way around. The other AI Act dates are listed further down under All AI Act dates.

When energy AI counts as high-risk

The decisive sentence sits in Annex III point 2, and it names the energy sector by name. High-risk is AI used as a safety component in the management and operation of critical infrastructure, or in the supply of water, gas, heating or electricity. No sub-clause. No room to argue about the industry. What remains open is what counts as a safety component in a given organisation, and on that the Commission published draft guidelines on 19 May 2026.

The draft turns this into two conditions that have to come together. First, the AI directly protects the physical integrity of the installation or the safety of people (paragraphs 184 to 189). If conventional grid protection with protection relays and fuses keeps working independently of the AI, it is not the AI that protects the installation. Second, the operator is designated a critical entity under the CER Directive (paragraphs 190 and 204). Whether a grid operator or heat supplier is one is decided in Germany by the KRITIS ordinance, which has not been issued yet. Two times yes means Annex III, and the classification of the individual system remains a question for your legal advisers.

A single system can become high-risk by two routes at once. If the AI sits inside a machine under the Machinery Regulation or in pressure equipment, Annex I applies on top. Both classifications carry their own duties, and one does not settle the other.

Concrete systems: what is covered and what is not

Now it gets practical. The dividing line does not run between departments, it runs between a direct protective function and everything else. The examples follow the draft guidelines of 19 May 2026 and only apply if the operator is a critical entity.

The draft explicitly lists load forecasts as not high-risk, including the imbalance forecast in 15-minute intervals, even though an error there can reach as far as a supply disruption. The chain to the damage is too long for it. Conversely, a human in the loop is no way out: anomaly detection that supports decisions on disconnections sits on the high-risk side. And everything depends on the operator's critical entity status: the same software can be high-risk at one grid operator and not at another. What that means for the AI inventory is covered in the piece on which AI at a utility is high-risk under Annex III.

The Article 6(3) filter

Not every AI that falls under Annex III is automatically high-risk. Article 6(3) carves out an exemption, and the Digital Omnibus sharpened it. For utilities this is the lever to narrow the number of genuinely affected systems honestly, without talking the duty away.

A system is not high-risk despite Annex III if it poses no significant risk to health, safety or fundamental rights, for instance because it performs a narrow procedural task or only prepares a human decision. For embedded AI the rule reads: if it merely optimises or supports a regulated product without its failure creating safety risks, the exemption applies.

One hard limit remains. As soon as the system profiles a natural person, the exemption is forfeited. It stays high-risk no matter how narrow the task. And whoever uses the exemption has to document the assessment before the system goes to market. The reasoning is part of the evidence, not a free-form note.

Which duties apply from 2 December 2027

Once a system is classified high-risk, the list of duties is long and concrete, and for AI in grid control systems it reaches into shift organisation. It is also the reason sixteen months are tight. Making a grid control system conformant is not a form, it is a project with data work, documentation and external assessment.

Add the technical documentation under Annex IV, the built-in logging under Article 12, and demonstrated accuracy, robustness and cybersecurity at the level the infrastructure demands. At the end sit a conformity assessment, registration in the EU database for high-risk AI, and clear instructions for use for the deployers. Most of this falls on the provider, usually the manufacturer. The utility as deployer must, under Article 26, use the system according to the instructions, assign human oversight to people with the competence and authority, monitor operation and keep the automatically generated logs for at least six months. That is a lot, and it sensibly does not start in November 2027.

AI Act meets NIS2 and EnWG

Here is the point many utilities underrate. The AI Act does not replace NIS2 or the EnWG security catalogues. It sits on top. The same grid control system then carries two sets of duties that overlap but do not coincide.

NIS2 and the KRITIS umbrella law govern the cybersecurity and resilience of operations through the BSI Act, in force since 6 December 2025, the same BSI that now requires post-quantum cryptography from 2030. The security catalogues under Section 5c EnWG, successor to Section 11(1a) and (1b), set a security level for grid control. The AI Act additionally governs the quality, verifiability and oversight of the AI system itself, supervised since 29 July 2026 by the Bundesnetzagentur as AI authority under the KI-MIG.

In practice: if you already built an asset inventory and a risk management system for NIS2, you can build on it, but you have to classify and document the AI systems separately. The same mindset is familiar to anyone who worked on OT security under IEC 62443. Duplicate work is avoided only by planning both frameworks together rather than one after the other.

Legacy systems and significant changes in design

Good news with fine print. AI already in use before 2 December 2027 is not initially covered by the new duties. But the grandfathering ends faster than it sounds, and public bodies face their own deadline anyway.

A high-risk system placed on the market or put into service before the cut-off only falls under the Annex III duties under Article 111(2) if it is subject to significant changes in its design from then on. A major model update or a new control function can be such a change and trigger the full duty. This is not the same as the substantial modification under Article 3 point 23, which decides the roles under Article 25.

The grandfathering is therefore no rest, it is a reason to check every planned change to a high-risk AI for its trigger effect first. That changes how decisions about updates are made.

All AI Act dates, relief and what is still open

The high-risk duties arrive late. Other parts of the AI Act have applied for a while, and some of them reach every utility that runs a customer chatbot.

• Since 2 February 2025: the prohibitions under Art. 5, such as social scoring and emotion recognition at work, with fines of up to EUR 35 million or 7 percent of worldwide annual turnover. Art. 4 on AI literacy has applied since the same day.
• Since 2 August 2025: obligations for providers of GPAI models.
• Since 29 July 2026: supervision by the Bundesnetzagentur under the KI-MIG.
• Since 2 August 2026: transparency obligations under Art. 50.
• 2 December 2026: machine-readable marking for generative systems placed on the market before 2 August 2026, plus the new prohibitions added by the Omnibus.
• 2 December 2027: Annex III. 2 August 2028: Annex I. 2 August 2030: existing systems of public bodies.

Since the Omnibus, Art. 4 only asks providers and deployers to take measures that promote the AI literacy of their staff. They don't have to guarantee a particular level, but they should still document what they do. What that looks like as an Art. 4 training concept, and what a utility needs for AI labelling under Art. 50, is covered separately.

Small mid-caps are new: companies that are not SMEs, have fewer than 750 employees and no more than EUR 150 million in annual turnover or EUR 129 million in balance sheet total. They get what used to be reserved for SMEs, namely simplified technical documentation, the lower fine ceiling and priority access to AI regulatory sandboxes. A mid-sized municipal utility should check whether it qualifies.

The duties themselves stay. Risk management, data governance, logging and human oversight apply to a utility with 400 employees exactly as they do to a group. What gets easier is the evidence, not the work.

And the interpretation? Not settled yet.

• The Commission's guidelines on high-risk classification have been a draft since 19 May 2026. The final version is announced for the end of 2026, and even then only the Court of Justice can interpret the regulation bindingly.
• Under the draft, whether grid AI can fall under Annex III depends on the operator being a critical entity. In Germany that is decided by the ordinance under the KRITIS umbrella act, which had not been issued by October 2026.
• CEN and CENELEC are working on further harmonised standards in committee JTC 21. The presumption of conformity only applies once their references are published in the Official Journal. The deadline stays where it is.

Anyone classifying grid AI today is classifying it against a draft. That is no reason to wait. It is a reason to write the classification down and look at it again once the final guidelines are out. The GDPR applies alongside, unchanged: a conformity assessment under the AI Act does not replace a data protection impact assessment.

What utilities should do now

The build-up to the end of 2027 works best in steps, and it starts not with technology but with visibility. Without a register of the AI in use, not a single duty can be met. The good news: anyone already steering AI tools in-house deliberately, for instance when building a sovereign AI stack, has done part of the work already.

Five steps to the deadline

1. Build an AI inventory

   Record every AI system, in-house and at vendors, that touches grid control, generation, metering or automation. Without this list every further assessment is patchwork. It takes a few days and decides everything after it.

2. Classify conservatively

   Check each system against Annex III point 2 and the Article 6(3) filter. When in doubt, classify higher and document the reasoning, especially for systems that can switch. The classification stays with the utility.

3. Review vendor contracts

   Clarify who demonstrates conformity, the AI vendor or you as the operator. Unassigned responsibility is the most common gap, and it surfaces in the audit, when it gets expensive.

4. Name an owner

   Assign responsibility for AI conformity to a role that connects IT, grid operations and legal. The AI Act is not a pure IT task, and no one fills the gap that sits between three departments.

5. Plan alongside NIS2

   Use the asset inventory and risk management from the NIS2 rollout as the base and add the AI-specific documentation. Two separate projects cost double, one shared project saves half.

What is assessed in the end is systems and their effect, not a single model. The model you run today may be replaced in two years. The classification and the duty to demonstrate stay. That is exactly why it pays to start now, and not once the standards are final.

Further Reading

Frequently Asked Questions