EU AI Act High-Risk AI Systems

The EU AI Act (Regulation 2024/1689) establishes a risk-based regulatory framework in which high-risk AI systems bear the most substantial compliance obligations. Article 6 defines two pathways to high-risk classification. First, under Article 6(1), AI systems that are safety components of products, or are themselves products, covered by the EU harmonised legislation listed in Annex I (including machinery, medical devices, toys, lifts, radio equipment, civil aviation, motor vehicles, and marine equipment) and that are required to undergo third-party conformity assessment under that legislation are classified as high-risk. Second, under Article 6(2), AI systems falling within the use cases enumerated in Annex III are classified as high-risk, regardless of whether they are embedded in a product.

Annex III enumerates eight areas of high-risk use: (1) biometrics, insofar as permitted under Union or national law; (2) critical infrastructure — management and operation of road traffic, and supply of water, gas, heating, and electricity; (3) education and vocational training — determining access to or influencing outcomes in education and training; (4) employment, workers management, and access to self-employment — recruitment, selection, contract decisions, monitoring, and evaluation of workers; (5) access to and enjoyment of essential private and public services and benefits — creditworthiness assessment, risk assessment in life and health insurance, evaluation of applications for public assistance and services, and emergency dispatch prioritisation; (6) law enforcement — individual risk assessments, polygraphs and similar tools, evaluation of evidence reliability, crime prediction, and profiling; (7) migration, asylum, and border control management; and (8) administration of justice and democratic processes.

Article 6(3) introduces an important qualification: an AI system listed in Annex III is not considered high-risk if it does not pose a significant risk of harm to health, safety, or fundamental rights. This exception applies where the AI system performs a narrow procedural task, improves the result of a previously completed human activity, detects decision-making patterns without replacing or influencing the human assessment, or performs a preparatory task to an assessment relevant to the Annex III use cases. Providers relying on this exception must document their assessment before placing the system on the market and must notify the relevant national competent authority. The exception is narrow and must be applied conservatively — its misapplication exposes the provider to the full high-risk compliance obligations retroactively.

Article 10 establishes data governance requirements that apply to training, validation, and testing datasets used for high-risk AI systems. The provision recognises that AI system quality is fundamentally dependent on data quality, and mandates that datasets be subject to appropriate data governance and management practices. These practices must address, at minimum: design choices for data collection and preparation, data collection processes and their origin, data preparation operations (annotation, labelling, cleaning, updating, enrichment, and aggregation), the formulation of relevant assumptions regarding what the data measures and represents, an assessment of the availability, quantity, and suitability of the datasets needed, and examination of possible biases that are likely to affect health, safety, or fundamental rights.

Training, validation, and testing datasets must be relevant, sufficiently representative, and, to the best extent possible, free of errors and complete in view of the intended purpose of the system. They must have the appropriate statistical properties, including with respect to the persons or groups of persons in relation to whom the high-risk AI system is intended to be used. Article 10(4) addresses bias explicitly: to the extent necessary for bias monitoring, detection, and correction, providers may process special categories of personal data under Article 9 of the GDPR, subject to appropriate safeguards including pseudonymisation and encryption. This provision creates a carefully bounded exemption that allows bias mitigation through the use of demographic data that would otherwise be off-limits.

For organisations operating under both the AI Act and the GDPR, data governance under Article 10 must be coordinated with GDPR obligations. The legal basis for processing training data must be established under Article 6 of the GDPR, purpose limitation under Article 5(1)(b) must be respected, data minimisation under Article 5(1)(c) must be applied, and data protection impact assessments under Article 35 must be conducted where appropriate. The AI Act's data quality requirements complement but do not replace GDPR data protection obligations — compliance with both frameworks is required, and tension between data quality (more data, more representative data) and data minimisation (less data, only necessary data) must be resolved through careful proportionality analysis documented in the risk management system.

Article 11 requires providers to draw up technical documentation before a high-risk AI system is placed on the market or put into service, and to keep that documentation up to date. The technical documentation must demonstrate compliance with the requirements of Chapter III, Section 2, and provide national competent authorities and notified bodies with all necessary information to assess compliance. Annex IV specifies the minimum content of technical documentation: a general description of the system, a detailed description of elements and development process, detailed information about monitoring, functioning, and control, a description of the risk management system, a description of changes made throughout the lifecycle, a list of harmonised standards or common specifications applied, a description of the conformity assessment procedure followed, and an EU declaration of conformity.

The conformity assessment procedure depends on the classification pathway. For AI systems classified as high-risk under Article 6(2) (Annex III use cases), providers generally follow the internal control procedure set out in Annex VI — a self-assessment that includes verifying the quality management system, examining the technical documentation, and verifying that the design and development process is consistent with the requirements. However, for biometric identification and categorisation systems, the conformity assessment requires the involvement of a notified body under Article 43(1). For AI systems classified as high-risk under Article 6(1) (Annex I product legislation), the conformity assessment follows the procedure applicable under the relevant sectoral legislation, integrating the AI Act requirements into the existing product conformity framework.

Upon successful conformity assessment, the provider issues an EU declaration of conformity (Article 47) and affixes the CE marking (Article 48). The declaration of conformity must contain specified information including the provider's identity, a statement that the declaration is issued under the sole responsibility of the provider, that the high-risk AI system complies with the AI Act, references to relevant harmonised standards or common specifications, and details of the notified body involved (where applicable). The technical documentation and declaration of conformity must be kept for a period of ten years after the high-risk AI system has been placed on the market or put into service. National market surveillance authorities may request access to this documentation at any time.

The EU AI Act entered into force on 1 August 2024, with a phased application schedule. Prohibited practices under Article 5 became applicable on 2 February 2025. AI literacy obligations under Article 4 and provisions relating to general-purpose AI models became applicable on 2 August 2025. The high-risk AI system requirements — the obligations described in this guide — become fully applicable on 2 August 2026 for newly placed systems. For high-risk AI systems that are also regulated as products under Annex I legislation (medical devices, machinery, etc.), the application date is 2 August 2027. This timeline gives providers and deployers a defined but finite window to achieve compliance.

Preparation should begin with an AI system inventory: catalogue all AI systems developed, deployed, or procured within your organisation, classify each against the Annex III use cases and the Article 6(3) exception criteria, and prioritise compliance efforts by risk classification. For systems classified as high-risk, establish the risk management system (Article 9), implement data governance practices (Article 10), develop technical documentation (Article 11, Annex IV), design for transparency and human oversight (Articles 13-14), and prepare the conformity assessment evidence. For organisations that are deployers rather than providers, focus on the deployer obligations under Article 26: use the system in accordance with instructions, ensure human oversight, monitor operation, maintain logs, and conduct a fundamental rights impact assessment where required.

The compliance effort for high-risk AI systems is substantial and should not be underestimated. Organisations that have invested in responsible AI governance frameworks — including risk assessment, bias testing, transparency documentation, and human review processes — will find the transition more manageable. Those starting from scratch should allocate dedicated resources, engage legal and compliance expertise with specific AI Act knowledge, and consider the AI Act's interaction with the GDPR (for data governance and individual rights), NIS2 (for cybersecurity requirements applicable to critical infrastructure AI), and sector-specific legislation (for regulated products and services). The AI Act is not a standalone regulation — it operates within the broader EU regulatory ecosystem, and compliance requires a holistic approach.