Governance
May 24, 2026
8 min read

EU AI Act for Agentic AI:
Technical Compliance Requirements

“The EU AI Act does not regulate models. It regulates systems — and autonomous agents that take consequential action are squarely in scope.”

Why Agentic AI Is in the EU AI Act's Crosshairs

The EU AI Act (Regulation 2024/1689), in force from August 2024 with phased application through 2027, takes a risk-based approach. It does not ban AI. It places technical and governance obligations on AI systems that operate in high-risk contexts — and it defines those contexts broadly enough to capture most autonomous agent deployments.

A key principle: the Act regulates the system, not the model. An agent that wraps GPT-4o with MCP tool access to HR systems, financial databases, or critical infrastructure is a regulated AI system regardless of which underlying model it uses.

High-Risk Classification Triggers

Annex III of the Act lists eight high-risk categories. Autonomous agents commonly intersect with:

Annex III §4
Employment & workforce
Agents that screen CVs, schedule interviews, or recommend terminations.
Annex III §5
Access to essential services
Agents processing loan applications, insurance underwriting, or credit scoring.
Annex III §6
Law enforcement
Agents used in evidence analysis, risk assessment, or investigative support.
Annex III §2
Critical infrastructure
Agents with tool access to energy management, water systems, or transport.
Annex III §3
Education
Agents that evaluate student work or make admissions recommendations.

If your agent operates in any of these categories and is deployed to EU users or by EU-based organizations, the high-risk obligations apply — regardless of where the system is hosted.

Key Technical Requirements

High-risk AI systems must satisfy requirements under Articles 9–17. The four most technically demanding for agentic deployments:

Article 13

Transparency and provision of information

Requirement: Users must be informed they are interacting with an AI system. The system must produce outputs that are interpretable by users and deployers.

Implementation: Every real agent action must be attributable and explainable. Tool calls should produce structured, reviewable records — not just an action and an outcome.

Article 14

Human oversight

Requirement: High-risk AI systems must be designed to allow human operators to intervene, override, or stop operation at any time. Oversight must be technically enforceable — not just a policy.

Implementation: Agents cannot be given unconditional tool authority. Every action that matters must be capable of interception before execution. An approved/blocked/escalated gate at the execution boundary satisfies this requirement directly.

Article 17

Quality management system

Requirement: Providers must implement a quality management system covering risk management, data governance, technical documentation, and post-market monitoring.

Implementation: Audit trails must be complete, tamper-evident, and retained. Cryptographically committed logs that cannot be retroactively modified satisfy the integrity requirements of Article 17.

Article 9

Risk management system

Requirement: A continuous, iterative risk management process must identify, analyze, and mitigate risks across the AI system lifecycle.

Implementation: Threat modeling of agent tool access surfaces (prompt injection, SSRF, privilege escalation) should be documented and reviewed regularly as part of the risk management record.

The Evidence Architecture That Satisfies All Three

Articles 13, 14, and 17 share a common infrastructure requirement: the system must be capable of producing a complete, accurate, tamper-evident record of what the agent decided, what it did, and why — at every step that matters.

This is what the VEX Protocol's Evidence Capsule architecture is designed to produce. Each capsule binds four pillars: authority, intent, identity, and witness — producing a record that satisfies the transparency, oversight, and audit requirements simultaneously.

For teams operating under DORA, SOC 2, or the NIS2 Directive alongside the AI Act, the same evidence architecture serves multiple compliance frameworks without duplicated instrumentation.

Design for compliance from the execution layer.

See how ProvnAI architecture maps to EU AI Act, DORA, and SOC 2 requirements.