AI Governance for Energy and Utilities: The Grid Is High-Risk, the Trading Desk Is Not

An energy company's AI sits on both sides of a hard line. On one side: AI that helps run the physical grid — voltage control, load balancing, automated protection, fault isolation. On the other: AI that runs the business — demand forecasting, wholesale trading, predictive maintenance scheduling, customer billing. Under Regulation (EU) 2024/1689 the first side can be high-risk and the second almost never is. A governance model that does not draw that line precisely either over-controls the trading desk or under-protects the grid.

"AI governance" is a management practice, not a term the EU AI Act defines. The Act imposes duties — classify by use, manage risk on high-risk systems, oversee, document — and a utility's governance model is the organisation that makes each duty land on a named owner. For energy, the distinctive challenge is that the high-risk systems live in the operational technology (OT) world — SCADA, EMS, DERMS, substation automation — while the governance and documentation discipline historically lives in IT and corporate risk. The model has to span both.

---

The high-risk hotspot: grid safety-component AI

The relevant Annex III heading for utilities is point 2 — critical infrastructure: AI systems intended to be used as safety components in the management and operation of critical digital infrastructure, road traffic, or the supply of water, gas, heating and electricity.

The two load-bearing words are safety component. An AI system is high-risk under Annex III point 2 when it performs a safety function whose failure could endanger life, health or property — for example, AI that controls or protects the physical supply: automated load-shedding, real-time voltage/frequency regulation, protection-relay coordination, or automated islanding decisions during a fault. If the AI's job is to keep the lights on safely, and its failure means they go off dangerously, it is the kind of system Annex III point 2 is aimed at.

Most AI in an energy company is not a safety component:

The distinction matters because energy firms routinely over-estimate their high-risk exposure. Forecasting and trading models — the AI an energy company invests most in — are commercial, not safety-critical. They sit under REMIT (wholesale market integrity) and, where personal data is involved, GDPR, but not the AI Act high-risk stack. The governance model should resist the temptation to gold-plate them into a conformity assessment they do not need.

---

The Article 6(3) filter: is the grid AI really high-risk?

Even an AI system that sits in an Annex III category is not automatically high-risk. Article 6(3) lets a provider conclude a system poses no significant risk of harm where it meets one of four conditions — a narrow procedural task; improving a previously completed human activity; detecting decision patterns without replacing or influencing a human assessment; or purely preparatory work. Only one condition is needed.

For grid AI this filter does real work. An AI that recommends a load-balancing action to a control-room operator who decides and acts may qualify as "preparatory" or "not replacing the human assessment" — pushing it out of high-risk. An AI that automatically trips a breaker or sheds load without human intervention is a genuine safety component and stays high-risk. The exemption is a documented judgment, not a loophole: the assessment must be written down, and the system still registered under Article 49. The governance model owns making — and recording — that call for each grid system.

---

Provider, deployer, or both

Most utilities are deployers under Article 26: they buy SCADA/EMS platforms and grid-automation modules from industrial vendors (the large automation OEMs) and operate them. The vendor that builds and ships the safety-component AI is the provider under Article 16 and carries the conformity-assessment burden.

The flip in Article 25 matters in OT: a utility that takes a vendor control module and substantially modifies its safety logic, or integrates several systems into a bespoke automation scheme it puts its own name to, can become a provider for that combined system. In a sector where in-house engineering teams customise vendor automation heavily, this is not a corner case — the governance model must test for it at change-control.

---

Governance that spans OT and IT

The defining feature of an energy AI governance model is that it must reach into the OT estate that corporate compliance functions rarely touch. Three design choices follow:

1. One register, both worlds. The use-case register must capture OT control-system AI alongside corporate AI. If the grid-automation team keeps its own list and corporate risk keeps another, the safety-component systems — the only high-risk ones — fall between them.
2. Control-engineering owns classification, second line reviews it. The person who can judge whether an AI module is a safety component is a control or protection engineer, not a compliance officer. The model federates classification to OT engineering and reviews it centrally — hub-and-spoke fits utilities well.
3. Map AI Act duties onto functional-safety practice. Energy OT already runs functional safety (IEC 61508 / IEC 61511) and change management. The AI Act's Article 9 risk management, Article 14 human oversight and Article 15 robustness map onto that existing safety discipline rather than replacing it — but the Article 11 / Annex IV documentation, the Article 43 conformity assessment and the Article 49 registration are new and need owners.

---

The NIS2 interplay

Energy is an "essential entity" sector under NIS2 (Directive (EU) 2022/2555), which imposes cybersecurity risk-management measures and incident reporting on grid operators. There is real overlap with the AI Act's Article 15 (accuracy, robustness, cybersecurity) for high-risk systems: a safety-component AI must be cyber-resilient under both regimes. Run the assessment once and let it satisfy both — the NIS2 risk-management measures for an OT control system and the Article 15 cybersecurity evidence for the same system are substantially the same evidence.

The incident-reporting duties differ, though, and the governance model must keep them distinct: NIS2 significant-incident reporting goes to the national CSIRT/competent authority; AI Act Article 73 serious-incident reporting goes to the market-surveillance authority. A single OT incident affecting a high-risk AI safety component may trigger both, on different timelines. Designing one incident-management workflow that routes to both authorities is the efficient answer; assuming one report covers both is the trap.

---

Worked example: a regional grid operator

Nordlys Netz GmbH — ~480 employees, ~€95M turnover, a regional distribution-system operator (DSO) in northern Germany — runs three AI systems worth governing: an automated load-shedding/protection module embedded in its substation automation, a renewables-and-demand forecasting model, and a customer-service chatbot.

Its governance model, hub-and-spoke spanning OT and IT:

• Accountable executive: the COO, who owns grid operations and functional safety, takes AI safety risk onto the operational risk map.
• Classification: the load-shedding/protection module is a safety component — high-risk under Annex III point 2 (its failure can endanger supply); Nordlys runs the Article 6(3) test and confirms it cannot be exempted because the action is automatic, not preparatory. The forecasting model is minimal. The chatbot is Article 50 limited.
• Role: Nordlys deploys a vendor automation module unmodified, so it is a deployer (Article 26); the automation OEM is the provider. Nordlys verifies the provider's Article 47 declaration and Article 13 instructions as part of OT vendor due diligence.
• The high-risk work, by 2 December 2027: as deployer, document Article 14 control-room oversight and keep operation logs (Article 26); an Article 27 FRIA is not required here (FRIA is for public-body and creditworthiness/insurance deployers, not grid operators). Nordlys also confirms the provider's conformity is current, aligns the Article 15 evidence with its NIS2 OT-security file, and wires AI Act Article 73 reporting into its existing NIS2 incident workflow.
• Live now: Article 4 AI literacy measures for control-room and engineering staff (in force since 2 February 2025); Article 5 prohibited-use screening at intake; the chatbot's Article 50 disclosure, required since 2 August 2026.

Nordlys's lesson: only one of its three systems is high-risk, it is a deployer not a provider for that system, and the heavy lifting is verification plus reusing its NIS2 and functional-safety evidence — not building a conformity dossier from scratch.

---

The deadline

For stand-alone high-risk systems, the original date was 2 August 2026. Under the Digital Omnibus (Commission proposal November 2025), the application date for stand-alone Annex III high-risk systems moved to 2 December 2027. Grid AI embedded as a safety component in a product covered by Annex I Section A law (radio, pressure or ATEX equipment, for example) follows the 2 August 2028 date for product-embedded high-risk AI. AI in machinery now sits in Annex I Section B: its requirements arrive through delegated acts under the Machinery Regulation (EU) 2023/1230, applying by 2 August 2028, rather than through the AI Act's conformity route directly.

The point to keep front of mind: the deferral is law — the Omnibus was published in the Official Journal as Regulation (EU) 2026/1744 on 24 July 2026 and has been in force since 27 July 2026, so the amended Regulation itself now reads 2 December 2027. Plan against 2 December 2027 for stand-alone Annex III. OT conformity evidence is slow to assemble, so this is not a reason to wait.

---

Penalties

Under Article 99: prohibited Article 5 practices — €35,000,000 or 7% of worldwide turnover (99(3)); high-risk obligations and Article 50 disclosure — €15,000,000 or 3% (99(4)); supplying incorrect or misleading information to an authority — €7,500,000 or 1% (99(5)). The third tier is 1%, never 1.5%. The Article 99(6) cap (lower of the percentage or the fixed sum) applies to entities meeting the SME definition. Since 27 July 2026, Article 99(6a) gives small mid-caps (SMCs) — a mid-sized operator like Nordlys can qualify — the same lower-of cap for the 99(4) and 99(5) tiers, though not for Article 5 fines.

---

How Confir helps

Confir steps a utility's intake through the Annex III points — including the "safety component" test in point 2 — in plain-English checklists, so a control engineer's answers, not a lawyer's guess, drive the result. It separates the one or two genuine safety-component systems from the forecasting and trading models that are minimal, derives the Article 16/26 role (including the Article 25 flip when OT teams modify vendor automation), and generates the Article 11 / Annex IV pack and Article 47 declaration. The logic is deterministic and rule-based — the same intake yields the same finding, with the rule that fired recorded for an auditor.

Related guides

• AI governance operating model
• AI governance for manufacturing
• Annex III critical-infrastructure AI
• Electricity-grid AI and the EU AI Act
• The Article 6(3) exemption explained
• Human oversight under Article 14
• EU AI Act deadlines and timeline