Tool · Guide
AI Act matrix for energy utilities: duties, deadlines and high-risk AI by role
What the AI Act, after the Digital Omnibus, asks of grid operators, metering point operators, suppliers and heat providers
- As of 17 Oct 2026
- 20 min read
- Timeline to 2028
Book an implementation check →
30 minutes on the project that's pressing for you right now. No sales loop.
Since 2 August 2026 the transparency obligations of the EU AI Act apply, and since 29 July 2026 supervision in Germany has an address: the Bundesnetzagentur. The high-risk obligations for AI in the grid only arrive on 2 December 2027. Whether they hit a utility at all depends on two conditions. This page sorts both by market role and date.
An overview for orientation, not legal advice and not a review of your individual case. Whether and from when an obligation applies to your company is for your legal department or counsel to confirm. As of 17 October 2026, subject to changes in the law.
- The Digital Omnibus, Regulation (EU) 2026/1744, has applied since 27 July 2026. It moves the high-risk obligations under Annex III to 2 December 2027 and under Annex I to 2 August 2028.
- The transparency obligations under Art. 50 have applied since 2 August 2026, for example to the customer service chatbot. Generative systems from before that date have until 2 December 2026 for machine-readable marking.
- AI in grid operation is not automatically high-risk. Under the Commission's draft guidelines of 19 May 2026 it has to be a safety component and run at a critical entity. Load forecasting does not qualify.
- Most utilities buy AI in and are therefore deployers. Their duties sit in Art. 26, plus Art. 4 and Art. 50.
- Under the KI-MIG, the Bundesnetzagentur supervises as market surveillance authority, single point of contact and complaints office.
- For managing directors, heads of grid and IT, and AI leads at energy utilities.
The four risk categories, applied to utilities
The EU AI Act sorts AI by use, not by technology. The same language model can fall under the transparency rules in customer service and be high-risk in pre-screening job applicants. An inventory therefore classifies each use on its own, not each product.
Status: Banned since 2 February 2025
Penalty: Up to EUR 35M or 7% of global turnover
Status: Annex III from 2 December 2027
Requirements: Providers need risk management, documentation and a conformity assessment. Deployers meet Art. 26.
Status: Applies since 2 August 2026
Requirements: Disclosure and marking under Art. 50
Status: No obligations beyond Art. 4
Requirements: Voluntary codes of conduct. AI literacy under Art. 4 still applies.
Timeline: what applies, what's coming
The Digital Omnibus rebuilt the timeline in July 2026. What was still a proposal in spring has been law since 27 July 2026. The table shows the position on 17 October 2026. Where a row links to the deadline calendar, you'll find the continuously checked entry with its source there.
| Date | Milestone | Status | What it means |
|---|---|---|---|
| 1 August 2024 | AI Regulation (EU) 2024/1689 enters into force | Applies | Starting point for every later deadline. |
| 2 February 2025 | Bans (Art. 5) and AI literacy (Art. 4) | Applies | Banned practices must be switched off. Art. 4 applies in the weaker Omnibus version: measures to support AI literacy, no guaranteed level per person. |
| 2 August 2025 | Obligations for providers of general-purpose AI models (GPAI) | Applies | Aimed at model providers. Using a finished model doesn't put you in scope here. |
| 19 May 2026 | Draft guidelines on high-risk classification (Art. 6(5)) | Draft | Consultation ran to 23 July 2026. The Commission has announced the final version for the end of 2026. |
| 27 July 2026 | Digital Omnibus, Regulation (EU) 2026/1744, in force | Applies | Moves Annex III and Annex I, weakens Art. 4, adds two bans. |
| 29 July 2026 | KI-MIG in force | Applies | The Bundesnetzagentur becomes market surveillance authority, single point of contact and complaints office. In the deadline calendar |
| 2 August 2026 | Transparency obligations (Art. 50) and enforcement | Applies | Chatbots and AI-generated content need a disclosure. National authorities enforce from now on, Art. 4 included, and the Commission gains its powers over GPAI providers. In the deadline calendar |
| 2 December 2026 | Marking for existing systems, new bans | Next deadline | Generative systems placed on the market before 2 August 2026 must mark their output in machine-readable form (Art. 50(2)). The transition for the new bans ends. In the deadline calendar |
| 2 August 2027 | Existing GPAI models, AI regulatory sandboxes | Future | Models on the market before August 2025 must meet the GPAI obligations. Each member state must have set up at least one AI regulatory sandbox by then. |
| 2 December 2027 | High-risk under Annex III | Future | Covers AI as a safety component in critical infrastructure and AI in HR. In the deadline calendar |
| 2 August 2028 | High-risk under Annex I | Future | AI in products with their own product law, such as machinery. |
Already applies: bans, AI literacy, GPAI
Since 2025. Without an AI inventory you can show neither that no banned practice is running nor that staff are being trained.
Since August 2026: transparency and supervision
The notice in the chatbot, the marking of generated content, and an authority that's allowed to ask.
By December 2027: classify high-risk
A little over 13 months sounds like plenty. For an inventory, a classification of every use, new clauses in supplier contracts and the deployer duties under Art. 26, it's tight. Especially as the KRITIS ordinance, on which the second high-risk condition depends, is still missing.
High-risk in the grid: two conditions, not a blanket verdict
Annex III point 2 covers AI intended as a safety component in the management and operation of electricity, gas, heat or water supply. For a long time the industry read that as: AI in the grid is high-risk. So did an earlier version of this page. The Commission's draft guidelines of 19 May 2026 draw the line much more tightly, and under them both conditions must be met.
The AI directly protects the physical integrity of the installation. It detects hazardous states, prevents start-up, limits or stops. AI that only informs, forecasts or optimises is not a safety component. Neither is pure cybersecurity AI (recital 55).
The operator is identified as a critical entity under the CER Directive, in Germany via the KRITIS umbrella act. The act has applied since 17 March 2026. Which installations qualify is set by the KRITIS ordinance, and in May 2026 that was only a ministry draft. The same software can therefore be high-risk at one utility and not at the next.
Perimeter protection of electricity infrastructure with cameras, radar or drone defence. Anomaly detection that supports decisions on load distribution, grid stability or shutdowns.
Load forecasting for grid optimisation. Predictive maintenance on gas pipelines whose safety systems keep running independently. Fault early warning to the control room, where a person decides. Cyber monitoring of grid communication.
The draft is neither final nor binding. It is, though, the most precise template there is, and an inspecting authority is unlikely to ignore it. What stays open is AI that switches by itself without an independent protection layer beside it. The examples in detail are in our article which AI at a utility is really high-risk.
What deployers of high-risk AI have to do
If a utility buys a high-risk system and uses it unchanged, it is the deployer. The conformity assessment then sits with the provider. From 2 December 2027 deployers must:
- use the system in line with the provider's instructions for use (Art. 26(1)),
- assign human oversight to people with the competence, training and authority for it (Art. 26(2)). Making oversight technically possible is the provider's job under Art. 14,
- keep input data relevant and representative, as far as they control it (Art. 26(4)),
- monitor operation, inform the provider and market surveillance of risks, and report serious incidents (Art. 26(5)),
- keep automatically generated logs for at least six months (Art. 26(6)),
- inform workers' representatives and affected staff before use in the workplace (Art. 26(7)),
- and, where the system talks to people or generates content, provide transparency under Art. 50.
A fundamental rights impact assessment under Art. 27 is not required for AI under Annex III point 2. These systems are registered at national level, not in the public EU database (Art. 49(5)).
Four steps to a classification
What is still open
- Guidelines: the draft of 19 May 2026 may still change before the final version.
- KRITIS ordinance: without it, the second condition can't be checked with certainty.
- Standards: the harmonised standards from CEN-CENELEC JTC 21 are only partly finished. Drafts on risk management, logging and human oversight are still in progress.
- Data protection: the GDPR applies alongside as soon as personal data goes into a model, such as load profiles of individual households.
Where AI meets the AI Act, by market role
A German multi-utility is often all of these at once: grid operator, metering point operator, supplier and heat provider. The duties follow the use, so it pays to look role by role. These are typical cases, not your individual case.
This is where the high-risk question lives. Perimeter protection at substations and anomaly detection that supports switching decisions can fall under Annex III point 2 if the operator is a critical entity. The draft excludes forecasting and information systems as long as a person decides. Open: AI that controls by itself, such as automatic dimming under Section 14a EnWG without an independent protection layer.
Plausibility checks, substitute values, photo checks of meter installation: not high-risk under the draft, because there's no protective function. Art. 4 for staff matters more, and the GDPR once household load profiles go into a model.
The customer service chatbot has been the most common duty since 2 August 2026: whoever writes to it must be able to tell it's an AI. Generated text and images in campaigns may need marking, depending on the case. Emotion recognition of call centre staff is banned; applied to callers, it counts as high-risk under Annex III point 1.
Annex III point 2 names gas and heat explicitly. Predictive maintenance on gas pipelines is not high-risk under the draft if the safety systems keep running independently. Nor is an inspection robot in a CHP plant that only detects and never intervenes. AI that stops a plant by itself when readings look wrong may be a different matter.
Applicant pre-screening and performance assessment are high-risk under Annex III point 4, regardless of KRITIS status. The works council has a say here anyway, and in municipal in-house companies the staff council.
Provider or deployer: the question before all others
Most utilities buy AI in and are therefore deployers. Anyone who develops AI and puts it into service, even only in-house, is a provider. The gap in obligations is large.
Core duties for high-risk AI: risk management and data governance (Art. 9, 10), technical documentation and logging (Art. 11, 12), building in human oversight (Art. 14), conformity assessment, CE marking and registration. Plus a quality management system and post-market monitoring.
Core duties for high-risk AI (Art. 26): use according to the instructions, human oversight by trained staff, logs kept for at least six months, informing the works council and the people affected. Art. 4 and Art. 50 apply to every AI system, not just high-risk ones.
Anyone who offers a high-risk system under their own name, substantially modifies it, or changes its intended purpose so that it becomes high-risk is treated as the provider. A case worth checking rather than assuming: a bought-in model retrained on your own grid data and used in the control room.
Applies since 2 February 2025, in a weaker form since the Omnibus. Providers and deployers take measures to support the AI literacy of their staff and of others operating AI on their behalf. They no longer have to ensure a particular level per person. Training records are still the simplest evidence when the authority asks.
General-purpose AI models (GPAI): rarely your duty, often your tool
Large language models, the kind behind the well-known chat assistants, are general-purpose AI models. The GPAI obligations have applied since 2 August 2025, and they fall on the model providers.
What model providers must deliver
- Documentation: technical documentation and information for downstream providers (Art. 53)
- Copyright: a policy to comply with EU copyright law
- Training data: a public summary of the training content
- Systemic risk: for models trained with more than 10²⁵ FLOP, evaluation, risk mitigation, serious incident reporting and cybersecurity (Art. 55)
The Commission published the GPAI Code of Practice on 10 July 2025, with chapters on transparency, copyright and safety and security. Since 2 August 2026 it can enforce the obligations. Models on the market before August 2025 have until 2 August 2027.
A utility builds on these models, in a customer service bot or document search, and is the deployer for that use. Under the Commission's GPAI guidelines of July 2025, you only become a model provider yourself if a modification uses more than a third of the original training compute. Hardly any utility gets there.
Penalty framework
The fines are set out in Art. 99. The higher of the fixed amount and the turnover share applies. For SMEs, and since the Omnibus also for small mid-caps, the lower one applies.
Penalty tiers
- Banned practices (Art. 5): up to EUR 35M or 7% of global turnover
- Provider, deployer and transparency obligations: up to EUR 15M or 3% of global turnover
- GPAI providers: up to EUR 15M or 3%, imposed by the Commission (Art. 101)
- Incorrect or incomplete information to authorities: up to EUR 7.5M or 1% of global turnover
The maximums are ceilings, not price lists. The last risk is the closer one: answering the Bundesnetzagentur wrongly, because nobody in-house has an overview of the systems in use, creates an offence of its own.
Germany: KI-MIG and the Bundesnetzagentur
Since 29 July 2026 Germany's act on market surveillance and innovation promotion for artificial intelligence, KI-MIG for short, has applied. It adds no new substantive obligations. It adds an authority that's allowed to ask.
What the KI-MIG sets out
- Bundesnetzagentur (BNetzA) as market surveillance authority: among other things for AI in critical infrastructure and HR, plus single point of contact and complaints office.
- Coordination: the BNetzA supports the other market surveillance authorities as a competence centre. In fully harmonised product areas the existing authorities stay responsible.
- AI service desk: with an online tool for a first estimate of the risk class, pointers to the European guidelines and to building AI literacy.
- AI regulatory sandbox: development and testing in a controlled environment, mainly for SMEs.
What the authority can actually do is covered in our article KI-MIG: supervision of your AI now has an address.
What applies alongside
- NIS2 implementation act, in force since 6 December 2025: AI systems are covered by general risk management under the BSIG. There is no AI-specific section.
- KRITIS umbrella act, since 17 March 2026: decides KRITIS status and with it the second high-risk condition.
- Data protection: Germany's data protection conference (DSK) has published guidance on AI (May 2024) and on generative AI with RAG (October 2025).
- Co-determination: Sections 80(3), 90(1) no. 3 and 95(2a) of the Works Constitution Act name AI explicitly; Section 87(1) no. 6 applies to systems that can monitor behaviour or performance. Municipal in-house companies and public-law institutions fall under state staff representation law.
AI regulatory sandboxes
Requirement: each member state must set up at least one AI regulatory sandbox. The Omnibus pushed the deadline back a year to 2 August 2027. In Germany, the Bundesnetzagentur is setting one up.
- Development, training, testing and validation with the authority alongside
- Free of charge for SMEs as a rule (Art. 58)
- Since the Omnibus, real-world testing of certain high-risk systems also outside the sandbox (Art. 60a)
Liability for damage to third parties remains in place during the test (Art. 57). For a grid AI project on the edge of high-risk, the route can still pay off: the classification gets settled with the authority, not afterwards.
What pays off now
There's time until December 2027. Three things pay off whatever the final guidelines look like.
An AI inventory with role, classification and reasoning for each use. Without it, no request from the Bundesnetzagentur can be answered cleanly.
Providers of high-risk AI must give deployers instructions for use and details on logging and human oversight. That belongs in the tender and the contract, not in acceptance testing.
There is no general "EU AI Act compliant" seal. High-risk systems carry the CE marking after their conformity assessment, and anyone advertising compliance needs evidence for it. A documented risk classification builds more trust than any statement of intent.
Sources and further reading
FAQ
The bans and the AI literacy duty (Art. 4) have applied since 2 February 2025, the obligations for providers of general-purpose AI models since 2 August 2025. Since 2 August 2026 the transparency obligations under Art. 50 apply and national authorities enforce, in Germany the Bundesnetzagentur. The high-risk obligations under Annex III follow on 2 December 2027, those under Annex I on 2 August 2028.
No. Under the Commission's draft guidelines of 19 May 2026, AI only falls under Annex III point 2 if, as a safety component, it directly protects the physical integrity of the installation and its operator is identified as a critical entity. The draft excludes load forecasting, predictive maintenance alongside independent safety systems and pure cyber monitoring. It includes perimeter protection and anomaly detection that supports switching decisions. The draft is not final yet, and the German KRITIS ordinance that determines critical status is still missing.
Regulation (EU) 2026/1744 has applied since 27 July 2026. It moves the high-risk obligations under Annex III to 2 December 2027 and under Annex I to 2 August 2028. It weakens the AI literacy duty, gives existing generative systems until 2 December 2026 for machine-readable marking and moves the deadline for national AI regulatory sandboxes to 2 August 2027. Newly banned is AI that generates non-consensual intimate content or child sexual abuse material.
The Bundesnetzagentur. Under the KI-MIG, in force since 29 July 2026, it is market surveillance authority, single point of contact and complaints office, among other things for AI in critical infrastructure and HR. Its AI service desk offers an online tool for a first estimate of the risk class.
For every AI system: measures to support staff AI literacy (Art. 4) and, where relevant, transparency under Art. 50, for example for a chatbot. For high-risk systems, from 2 December 2027 also the duties under Art. 26: use according to the instructions, human oversight by trained staff, monitoring operation, keeping logs for at least six months and informing workers' representatives before use in the workplace.
Up to EUR 35M or 7% of global annual turnover for banned practices, up to EUR 15M or 3% for breaches of other obligations and up to EUR 7.5M or 1% for incorrect information to authorities (Art. 99). The higher amount applies, for SMEs the lower one.
With an inventory of every AI use, including inside specialist software. For each use, settle the role, provider or deployer, and record the classification with its reasoning. Treat KRITIS status as an open assumption until the KRITIS ordinance is issued. Whether an obligation applies in the individual case is for the legal department or counsel to confirm.
From obligation to governance
This page shows what the EU AI Act asks of energy utilities. How that turns into AI governance in-house, with an inventory, classification, sign-off paths and clear owners, is described in our service AI governance for energy utilities. Which AI applications are realistic for each market role is shown in the AI use case map for utilities.
All articles on the subject are collected in the topic field AI in the energy industry. The dates are in the deadline calendar, and the route from pilot to regular operation is under Prototype and AI.
Let's collaborate
One contact, a team behind it
Thirty minutes on what you're working on, whether grid, metering, market processes, IT architecture or AI. Afterwards you'll know how we'd approach it and who would sit at the table.
Energy and measurement technology since 2010 · innobu GmbH since 2014 · Based in Schleswig-Holstein, Germany
Whoever pitches also delivers: whoever leads the first call stays your contact right through to the steering committee, with established partners in the project. Direct: info@innobu.com