EU AI Act — Art. 53 / high-risk
The EU AI Act is the first comprehensive AI-specific regulation from a major regulator. It sets binding obligations for 'high-risk' AI systems, including AI agents used in regulated domains such as healthcare, employment, and critical infrastructure.
The Act (Regulation (EU) 2024/1689) classifies AI systems by risk level. High-risk systems face obligations under Article 53 and related provisions, including: a risk-management system across the AI system's lifecycle, technical documentation, logging capable of enabling traceability, transparency and instructions for use, human-oversight measures, and accuracy/robustness/cybersecurity evidence.
For an AI agent, 'conformance' means a verification pack checks coverage of these obligations directly — not a general capability score. AgentGrading's conformance packs anchor each check to a specific article so a pass or fail traces back to the regulatory text, not to an internal opinion of what 'good' looks like.
Timing note: the high-risk obligations were deferred by the Digital Omnibus on AI, signed 8 July 2026 — stand-alone Annex III systems to 2 December 2027, and high-risk AI embedded in Annex I regulated products to 2 August 2028, subject to publication in the Official Journal. General-purpose AI model obligations became enforceable by the AI Office on 2 August 2026, and most Article 50 transparency duties remain anchored to that date.
This page summarizes the obligations relevant to agent evaluation; it is not legal advice. Consult the official text and qualified counsel for compliance decisions.
Packs anchored to this standard
- Medical RAG — groundedness & abstention
- Financial reporting RAG — groundedness & abstention
- Insurance policy RAG — groundedness & abstention
- Public-sector benefits RAG — groundedness & abstention
- EU AI Act — high-risk conformance
Guides on this standard
- The EU AI Act on 2 August 2026: what actually applies — What took effect, and what the Digital Omnibus deferred to December 2027.
- How to map an eval to the EU AI Act — Anchoring each check to a specific article, so a pass traces back to the text.