EU AI Act 2025: How to Classify and Comply with High-Risk AI Systems

The EU AI Act (Regulation (EU) 2024/1689) divides AI systems into four risk tiers. High-risk systems — defined by Annex III or their role as safety components in regulated products — face the most demanding obligations: conformity assessments, mandatory technical documentation, and ongoing post-market monitoring. Providers that deploy these systems in the EU market must be ready to comply by August 2026, regardless of where they are headquartered.

Which AI systems fall under Annex III of the EU AI Act?

Annex III is the operative list most organisations should work through first. The eight categories listed above are not exhaustive technology types — they are defined by use case and consequence. An AI system that recommends job candidates is high-risk; one that schedules internal meetings is not. The distinction is whether the output materially affects a natural person's access to education, employment, essential services, or fundamental rights.

The Commission retains the power to update Annex III by delegated act. The EU AI Office published updated guidance in early 2025 clarifying that general-purpose AI models (GPAI) integrated into Annex III applications inherit the high-risk classification of the downstream application — the GPAI provider bears separate transparency and documentation obligations, but the deployer remains responsible for conformity of the complete system.

One important exclusion: Article 6(3) creates a self-assessment exemption for Annex III systems that are "narrow procedural tasks", do not produce profiles of individuals, do not replace human assessment, and whose output is easily reversible. Providers claiming this exemption must document their reasoning — it is not a blanket get-out.

What obligations apply to high-risk AI providers?

Chapter III, Section 2 of the AI Act sets out nine categories of obligation for providers of high-risk systems. These apply from the point of placing a system on the EU market or putting it into service, not from when enforcement begins.

• Risk management system (Art. 9): A continuous, iterative process spanning the full lifecycle — design through decommissioning. Must identify known and reasonably foreseeable risks, estimate their probability and severity, and document mitigation measures. Not a one-off audit.
• Data governance (Art. 10): Training, validation, and test datasets must meet quality criteria for relevance, representativeness, and freedom from bias. Providers must document data origins, processing steps, and any known limitations.
• Technical documentation (Art. 11 + Annex IV): A detailed dossier covering system architecture, training methodology, performance metrics, limitations, and foreseeable misuse scenarios. Must be kept up to date and produced on request for market surveillance authorities.
• Record-keeping and logging (Art. 12): Automatic logging of events sufficient to trace the system's operation across its lifespan — essential for post-incident investigation.
• Transparency towards deployers (Art. 13): Providers must supply deployers with clear instructions for use, including intended purpose, known performance limitations, maintenance requirements, and human oversight mechanisms.
• Human oversight (Art. 14): High-risk systems must be designed so that a natural person can effectively monitor, understand, intervene, override, or halt the system. This cannot be a nominal capability — it must be practically achievable.
• Accuracy, robustness, and cybersecurity (Art. 15): Systems must achieve declared performance levels under reasonably foreseeable conditions, including adversarial inputs.
• Quality management system (Art. 17): Providers with more than 10 employees must establish a documented QMS covering development, monitoring, and corrective action procedures.
• Post-market monitoring (Art. 72): Active collection and analysis of data on system performance in real-world deployment. Serious incidents must be reported to national market surveillance authorities within defined timeframes.

What is the conformity assessment process for high-risk AI?

The conformity assessment is the formal process by which providers demonstrate their high-risk AI system meets the Act's requirements before market placement. The outcome — an EU Declaration of Conformity — is functionally equivalent to CE marking for the AI system.

For most Annex III systems, the AI Act permits internal conformity assessment (Annex VI procedure): the provider conducts and documents its own assessment, registers the system in the EU database for high-risk AI (operational from August 2026), and issues the Declaration of Conformity. Third-party involvement is mandatory only for:

• Biometric identification systems (except verification with explicit user consent)
• AI systems that are safety components of products already requiring third-party assessment under sectoral legislation (medical devices, machinery, etc.)

Which sectors face the strictest AI Act enforcement?

Enforcement is risk-proportionate. Financial services, healthcare, and employment are the three sectors expected to attract the earliest enforcement attention. Credit-scoring AI systems fall squarely within Annex III category 5. HR AI tools for CV screening and performance monitoring sit in category 4. Both sectors have existing regulatory frameworks (EBA guidelines, GDPR profiling restrictions) that will interact with AI Act obligations, creating overlapping compliance workstreams from mid-2026.

How does the AI Act affect property management AI tools?

Property management AI sits, in most configurations, below the high-risk threshold. A tool that answers residents' questions about building rules, generates meeting minutes, or flags maintenance requests does not fall within any of the eight Annex III categories. It does not determine access to housing, employment, or financial services; it does not produce biometric data; it does not assist law enforcement.

This places standard property management AI under the limited risk category (Art. 50), which carries two main obligations: (1) users must be informed they are interacting with an AI system, and (2) the system must not be designed to impersonate natural persons in ways that could deceive users to their detriment. Both are straightforward to implement.

The classification changes if the AI system begins making or substantively influencing binding decisions — for example, AI that produces creditworthiness assessments of tenants for rental applications, or AI used in employment decisions about property management staff. These use cases push into Annex III categories 4 and 5 respectively, triggering the full high-risk compliance programme.

Frequently Asked Questions

What is the EU AI Act compliance timeline for high-risk AI systems?

The AI Act entered into force on 1 August 2024. Prohibited AI practices became unlawful on 2 February 2025. High-risk AI obligations (Chapters III and IV) apply from 2 August 2026. GPAI model obligations applied from 2 August 2025. The EU AI database — where high-risk systems must be registered — was due to be operational by August 2026.

How do GDPR obligations overlap with the EU AI Act?

The overlap is substantial and deliberate. GDPR's data minimisation, purpose limitation, and accuracy principles align closely with the AI Act's data governance requirements (Art. 10). The AI Act's transparency obligations reinforce GDPR's right to explanation for automated decisions (Art. 22 GDPR). The simplest approach is to extend your existing DPIA process to incorporate AI Act classification and risk management documentation simultaneously.

What are the fines for non-compliance with EU AI Act high-risk obligations?

The fine structure is tiered. Violations relating to high-risk AI systems carry fines of up to €15 million or 3% of global annual turnover, whichever is higher. Violations of prohibited practices (Art. 5) carry fines up to €35 million or 7% of global annual turnover. Providing false or misleading information to authorities can result in fines up to €7.5 million or 1.5% of turnover.

Is there a CE marking equivalent under the EU AI Act?

For standalone Annex III AI systems, the equivalent is the EU Declaration of Conformity, issued after completing the conformity assessment under Annex VI. This Declaration must be kept for 10 years after market placement and made available to authorities on request. Registration in the EU AI database functions as the publicly accessible record of compliance.

Does the EU AI Act apply to UK-based organisations post-Brexit?

Yes, in the same way the GDPR applies extraterritorially. If a UK-based organisation places a high-risk AI system on the EU market, or uses a high-risk AI system whose outputs affect individuals located in the EU, the AI Act applies. UK organisations without an EU establishment must appoint an EU-based authorised representative.

What is a fundamental rights impact assessment and when is it required?

Article 27 requires deployers of certain high-risk AI systems — specifically public bodies and private operators providing public services — to conduct a fundamental rights impact assessment (FRIA) before deploying. The FRIA requires the deployer to assess the potential impact on privacy, non-discrimination, freedom of expression, access to justice, and other fundamental rights. The assessment must be registered in the EU AI database.