AI systems used as safety components in the management and operation of the supply of water, gas, heating and electricity, and of critical digital infrastructure, are listed in Annex III point 2 and are candidates for high-risk classification under Article 6, subject to the Article 6 conditions. Systems embedded in regulated products in those sectors can also follow the Annex I route. Registration for point 2 systems is national, the Annex III rules apply from 2 December 2027 and the Annex I rules from 2 August 2028, and sector resilience, safety and incident duties apply alongside the AI Act. This guide does not decide any operational system and does not give engineering advice on network safety or resilience.
Which energy and water functions are listed?
Annex III point 2 lists AI systems intended to be used as safety components in the management and operation of the supply of water, gas, heating and electricity. That covers systems that control, protect or manage the network in ways that fulfil a safety function, such as controlling pressure, flow, voltage or temperature, detecting faults and isolating them, or managing load in a way that prevents a hazardous failure. A system that only forecasts demand, only optimises cost, only handles billing or only reports on network performance, without fulfilling a safety function, is assessed differently, and the distinction must be recorded from the system's function and failure modes. The record should describe the network, the safety function, the management or operation role, and the consequence if the system fails or malfunctions, with sector safety expertise. A forecasting model that later gains a control function has changed its function, and its classification must be reassessed before the change goes live.
What is critical digital infrastructure?
Critical digital infrastructure, in the Annex III context, means the digital infrastructure whose disruption would have a significant impact on the provision of essential services or on the functioning of other critical infrastructure. That can include data centres, cloud infrastructure, communication networks and the digital control systems on which energy, water, transport and other sectors depend. An AI system used as a safety component in the management and operation of such infrastructure is within the Annex III listing, subject to the Article 6 conditions, and its failure modes must be assessed in terms of the dependent services, not only the digital system itself. The record should describe the infrastructure, the services that depend on it, the safety function and the cascading consequences of failure. A system that manages a single application, without a safety function for infrastructure on which others depend, is assessed differently, and the distinction should be recorded with the reasoning and the sector analysis that supports it.
For the broader regulatory framework, see finance, insurance and critical infrastructure.
How do resilience and incident duties interact?
Energy, water and digital infrastructure operators are subject to sector resilience, cybersecurity and incident reporting duties that sit alongside the AI Act's post-market monitoring, serious incident reporting and cybersecurity requirements. A failure, a cyber incident or a serious incident under the AI Act can also be a reportable incident under sector law, with a different recipient, timescale and content, and the operator should diarise each duty separately rather than assuming one report satisfies all. The AI Act's serious incident reporting under Article 73, to the market surveillance authority, and the sector incident reporting, to the sector regulator or computer security incident response team, should be coordinated through a single incident log, with a named owner, a timeline and a record of each notification made. The operator should also plan for resilience testing, backup and recovery, and for the human oversight of automated control actions, since a control system that cannot be overridden or recovered safely is a resilience risk as well as a compliance risk.
What data and robustness questions arise?
Infrastructure control systems depend on sensor data, operational data and model inputs whose quality, timeliness and integrity directly affect safety. The AI Act's data governance and accuracy, robustness and cybersecurity requirements therefore have operational bite: the provider and deployer must address data quality, representativeness, bias in the operational context, adversarial robustness, data poisoning, and the handling of out-of-distribution inputs and sensor failures. The system should be tested against realistic operating conditions, including degraded data, communication failures and unexpected demand, and the testing evidence should be specific to the network and environment concerned. The deployer should monitor performance in operation, including drift, error rates and near misses, and should feed findings back to the provider through post-market monitoring. A system validated only on historical or simulated data, without deployment-specific testing and monitoring, does not support a safety-critical deployment decision.
How should an infrastructure deployment be evidenced?
The deployment file should record the infrastructure and safety function, the Article 5 screen, the Article 6 route and reasoning, the national registration record, the sector safety and resilience assessments, the technical documentation, the testing and validation evidence, the oversight and override design, the incident and recovery plans, and the monitoring and change records, all linked by the system version identifier. The file should show how the AI Act requirements are addressed within the sector safety file, and where sector requirements add duties beyond the AI Act. Changes to the system, its data, its operating environment or its safety function should trigger a joint review, with the change, the reassessment and the approval recorded. The file should be available to the market surveillance authority and the sector regulator, in the form each requires, without describing the system differently in each forum. A file that is assembled only after an incident will be incomplete, and its gaps will be visible to the authorities that investigate.
Continue with critical infrastructure safety components.
What does this guide not decide?
This guide does not decide whether any energy, water or digital infrastructure system is a safety component, is high-risk, or follows the Annex I or Annex III route, and it does not give engineering advice on network design, safety or resilience. Those questions require the system's function, failure modes, sector legislation and safety analysis to be examined together, with sector expertise and qualified advice. The guide also does not describe any Member State's national registration procedure or sector incident reporting rules, which must be confirmed with the authorities concerned. Where an infrastructure system with a safety function is to be deployed, confirm the route, registration, conformity and sector safety position with the competent authorities before the system is put into service.
Frequently asked questions
Is a demand forecasting model high-risk?
It depends on its function. A forecast that only informs planning is assessed differently from a system used as a safety component in managing or operating supply. If the function changes to control, the classification must be reassessed.
Does one incident report satisfy AI Act and sector duties?
No. Article 73 reporting to the market surveillance authority and sector incident reporting can have different recipients, timescales and contents. Each duty should be diarised separately and coordinated through a single incident log.
Where are point 2 systems registered?
At national level under Article 49(5), not in the EU database. The process and authority must be confirmed in each Member State concerned.
Continue through this AI Act cluster
This article is an informational overview, not a legal determination. Check the current legal text and official implementation guidance for your system and role.
Consolidated Regulation (EU) 2024/1689, as amended ↗AI Act Service Desk implementation timeline ↗