Is Electricity Grid AI High-Risk Under the EU AI Act? Annex III Point 2

It depends on what the AI does to the grid. If the system is a safety component in managing or operating the supply of electricity — the part that keeps the lights on and the network stable — it is high-risk under Annex III point 2 of the EU AI Act, Regulation (EU) 2024/1689. If it forecasts demand so a trading desk can place better bids, it is almost certainly minimal-risk. Same utility, same data, two different tiers. The deciding word is "safety component," and it is the whole test.

What Annex III point 2 actually says

Point 2 makes high-risk the AI intended to be used as a safety component in the management and operation of critical digital infrastructure, road traffic, or the supply of water, gas, heating and electricity. Electricity supply is named explicitly. But the category is not "any AI at an energy company" — it is AI whose failure or malfunctioning would endanger the life and health of persons or cause significant disruption to the supply. That qualifier is doing the heavy lifting.

A safety component (the concept Annex III point 2 turns on) is defined in Article 3(14), which Regulation (EU) 2026/1744, the Digital Omnibus on AI, tightened from 27 July 2026: a component that fulfils a safety function — its intended purpose is to prevent or mitigate risks to the health and safety of persons or property — or whose failure or malfunctioning endangers health and safety. Pure optimisation or efficiency models do not qualify through the safety-function limb. In grid terms: real-time control that holds frequency and voltage within limits, protection systems that trip to prevent cascading failure, and automated balancing that keeps generation matched to load. If the model is wired into that loop, point 2 is in play.

The line that decides it

Ask one question: if this AI gave a wrong answer at the worst moment, could it destabilise the grid or endanger people?

The trap is treating "grid AI" as one thing. A transmission system operator might run a high-risk balancing controller and a minimal-risk price-forecasting model side by side. Classify each system on its own function — never the company, never the dataset, never the vendor.

Forecasting for trading is the clearest non-safety case. A demand-forecast model that informs market bids optimises money, not network stability. If it is wrong, someone trades badly — nobody is endangered and the supply is not disrupted by the model itself. That is minimal-risk. The moment the same forecast is fed directly into an automated dispatch loop that physically balances the grid without a human in control, you are back in safety-component territory and re-classify.

What a high-risk grid system owes

Where a system is a safety component under point 2, the full high-risk regime applies to the provider and deployer:

• Risk management across the lifecycle (Article 9)
• Data governance for training, validation and testing data (Article 10)
• Technical documentation to Annex IV (Article 11) and event logging (Article 12)
• Transparency / instructions for use to the deployer (Article 13) and human oversight (Article 14)
• Accuracy, robustness and cybersecurity (Article 15) — the load-bearing requirement for grid control
• Conformity assessment via the Annex VI internal-control route (Article 43), the declaration of conformity (Article 47), CE marking (Article 48) and national-level registration (Article 49(5) — point-2 systems are excepted from the public EU database)
• Deployer duties under Article 26, including oversight and log retention

Two parallel regimes also bite for grid operators: NIS2 cybersecurity obligations and sectoral electricity-network codes. They run alongside the AI Act, not instead of it — the Article 15 cybersecurity requirement and NIS2 overlap heavily for a control-system AI.

Note the route: critical-infrastructure AI uses the Annex VI internal self-assessment under Article 43, not the notified-body route reserved for Annex III biometrics. A notified body is not generally required.

Worked example: a transmission operator's two models

Netzkraft AG, a regional transmission operator with about 240 employees and roughly €95M turnover, runs two AI systems.

The first forecasts day-ahead demand to inform its market position. It touches no control loop; if it is wrong, the trading desk simply hedges. Not a safety component → minimal-risk. Netzkraft documents the classification and moves on — no Article 9–15 stack required.

The second is an automated balancing controller that adjusts dispatch in near-real time to hold frequency. Its failure could destabilise the network — a textbook safety component under Annex III point 2 → high-risk. Netzkraft, as deployer (and its vendor as provider), scopes the full regime: risk management (Article 9), human oversight so an operator can intervene (Article 14), and heavy investment in accuracy, robustness and cybersecurity (Article 15) aligned with its NIS2 duties. Conformity goes through the Annex VI internal route (Article 43) and the system is registered at national level (Article 49(5)), not in the public EU database. One company, two systems, two tiers — exactly as the safety-component test predicts.

How Confir helps

Confir runs the safety-component test against each grid AI use, separating minimal-risk forecasting, trading and billing models from the high-risk control and balancing systems that sit in the safety loop — then assembles the provider/deployer obligations for the systems that qualify, with the Article 15 robustness-and-cybersecurity emphasis grid control demands and a cross-map to NIS2. The engine is deterministic and rule-based: the same intake produces the same classification every time.

Related guides

• Critical infrastructure AI: Annex III point 2
• EU AI Act for energy and utilities
• How Article 6 classifies high-risk AI
• Article 15 accuracy, robustness and cybersecurity
• What a safety component is
• How AI risk classification works