Business

AI Act

The EU regulation on artificial intelligence (EU 2024/1689): risk-based classification of AI systems, obligations and fines of up to €35M or 7% of global revenue.

Stèle épaisse en verre givré dressée sur un sol sombre, un sceau de cire bordeaux luisant en son centre, une petite balance de justice en laiton posée devant, sous un faisceau de lumière blanche
AI Act — Image générée par IA (Qwen-Image 2.1, ComfyUI local)

The AI Act (Regulation EU 2024/1689) is the European Union's regulation on artificial intelligence — the world's first horizontal legal framework governing AI. In force since August 1, 2024, it classifies AI systems by risk level and imposes obligations proportionate to that risk, with a phased rollout through August 2, 2026 (general applicability) and fines of up to €35 million or 7% of global revenue.

A risk-based approach

The AI Act doesn't regulate the technology itself but its uses, split into four levels:

Level Examples Regime
Unacceptable risk Social scoring, behavioral manipulation, certain biometric uses Prohibited (since February 2025)
High risk Hiring, credit scoring, biometrics, critical infrastructure Strict obligations (documentation, human oversight, robustness)
Limited risk Chatbots, consumer generative AI Transparency obligations (inform the user)
Minimal risk Spam filters, video games No specific obligation

On top of this sits a dedicated regime for general-purpose AI (GPAI) models — large foundation models such as GPT, Claude or Mistral — with stricter obligations for those that pose a systemic risk.

The AI Act shifts responsibility to the organization that deploys AI, not only to the one that builds it. In practice, a company that uses an AI agent to screen job applications or decide on a loan falls within the "high-risk" scope and must be able to demonstrate effective human oversight, traceable decisions and proper information for the people affected. That's exactly the governance logic AI agent deployments need to build in — knowing, for every automated action, who approves it and who is accountable.

Key dates

  • February 2025: the prohibitions (unacceptable-risk practices) take effect.
  • August 2, 2025: obligations for general-purpose AI (GPAI) models apply.
  • August 2, 2026: general applicability of the regulation, including a large share of high-risk systems.

This deadline makes compliance a project to plan ahead — mapping AI uses, classifying them by risk, putting oversight and documentation in place — rather than a reaction after an audit.

Standards and Specifications

Regulation (EU) 2024/1689 (AI Act)

Official text of the European regulation on AI, in force since August 1, 2024, the world's first horizontal legal framework for artificial intelligence

Frequently Asked Questions

Is my company affected by the AI Act if it doesn't build AI?

Yes, most likely. The AI Act distinguishes several roles: the provider (who develops an AI system or has one developed), the deployer (who uses an AI system in a professional context), the importer and the distributor. Most companies are deployers: as soon as you use an AI tool to hire, score customers, moderate content or make decisions that affect people, you carry obligations (transparency, human oversight, log retention). Not building the model doesn't exempt you from the obligations tied to using it.

What changes with the August 2, 2026 deadline?

The AI Act applies in stages. The prohibitions (social scoring, manipulation, certain biometric uses) have applied since February 2025. The obligations for general-purpose AI (GPAI) models since August 2, 2025. August 2, 2026 marks the general applicability of the regulation, notably for a large share of high-risk systems. It's the date that turns AI governance from a best practice into an enforceable legal obligation for most organizations.

What are the penalties for non-compliance?

Fines are tiered by severity. The highest ceiling applies to prohibited practices (Article 5): up to €35 million or 7% of total worldwide annual turnover, whichever is higher. Breaches of the obligations for high-risk systems and GPAI models fall under lower ceilings (€15M / 3% and €7.5M / 1.5% depending on the case). The order of magnitude is comparable to the GDPR, which makes it a board-level financial risk.

Resources and Documentation