RHETRA / Guides / EU AI Act

EU AI Act compliance, practically

The EU AI Act requires organisations running high-risk AI to operate a risk-management system, data governance, automatic logging, transparency, human oversight and cybersecurity — continuously, with evidence (Art. 9–15). Fines for high-risk non-compliance reach €15M or 3% of global turnover. The practical question is not "which documents do we need?" but "can we produce verifiable evidence of what our AI actually did?" This guide covers the timeline, the obligations, and the evidence-first way to meet them.

1 · The timeline that matters

• 1 August 2024 — the Act entered into force.
• 2 February 2025 — prohibited practices apply (social scoring, untargeted facial-image scraping, and others).
• 2 August 2025 — obligations for general-purpose AI models apply.
• 2 August 2026 — the bulk of high-risk obligations apply (Annex III systems: credit scoring, employment, critical infrastructure, law enforcement and more).
• 2 August 2027 — extended transition ends for high-risk AI embedded in already-regulated products (medical devices, machinery).

If your organisation is a bank, insurer, hospital operator or public authority using AI in decisions about people, the August 2026 date is the one on your desk. Building the logging and oversight infrastructure takes quarters, not weeks — which is why supervisors expect to see it running, not planned.

2 · What high-risk actually obliges you to do

Seven operational duties, all continuous:

• Art. 9 — Risk management: an iterated, documented process over the whole lifecycle, not a one-time assessment.
• Art. 10 — Data governance: training/validation/test data must be relevant, representative and examined for bias.
• Art. 11 — Technical documentation: current, complete, and available to authorities.
• Art. 12 — Logging: the system must automatically record events over its lifetime — this is where most architectures quietly fail, because logs that can be edited are records, not evidence.
• Art. 13 — Transparency: deployers must be able to understand and use the system's output correctly.
• Art. 14 — Human oversight: effective, designed-in intervention capability — not a rubber stamp at the end.
• Art. 15 — Accuracy, robustness, cybersecurity: resilience against errors and against attacks, including data and model poisoning.

3 · The real fine tiers

Art. 99 defines three tiers — often misquoted, so precisely:

• Up to €35M or 7% of global annual turnover — for prohibited practices (Art. 5) only.
• Up to €15M or 3% — for violations of the other obligations, including all the high-risk duties above.
• Up to €7.5M or 1% — for supplying incorrect or misleading information to authorities.

4 · Documents vs evidence — the compliance trap

The default corporate response to regulation is documents: policies, registers, PDFs. The AI Act's high-risk duties are largely operational — they concern what the system does at runtime. A policy PDF cannot show that a specific AI call on a specific date was checked, redacted, allowed or blocked. Editable logs cannot prove they weren't cleaned up afterwards.

Evidence has two properties documents lack: it is generated by the enforcing system itself, and it is tamper-evident. That is achievable today: an enforcement gateway in front of every AI call, plus cryptographic signatures over every decision. Post-quantum signatures (e.g. Dilithium5 / ML-DSA, NIST-standardised) keep that evidence durable even against future quantum attacks — relevant, because audit trails must stay verifiable for years.

5 · The loop: author → enforce → prove → export

• Author the policy machine-readably: what data classes may reach which models, what gets redacted, who approves exceptions.
• Enforce it inline: a gateway (OpenAI-compatible, so applications don't change) classifies every call and applies BLOCK / SANITIZE / ALLOW — this is Art. 12 logging and Art. 14 oversight, produced as a side effect of operation.
• Prove each decision: sign it, and attest the runtime itself (hardware attestation, explained here).
• Export the evidence: give the auditor verifiable records instead of screenshots.

This is the architecture RHETRA ships: the KRONOS gateway and the Governance Platform implement the loop end to end — on-premise or SaaS. The wider context — why verifiability is the core of sovereignty — is in our pillar guide: What is Sovereign AI?

This guide is general information on the Regulation (EU) 2024/1689, not legal advice.