The EU AI Act applies in four waves from February 2025 to August 2027. The first obligation is always classification: prohibited, high-risk, limited, minimal. Mis-classification is one of the most expensive regulatory risks. Marking a high-risk system as minimal carries fines up to 15 M EUR plus market withdrawal. Here is a pragmatic five-step workflow.

Step 1: Build an AI inventory

Classification starts with an inventory. Which AI systems are in use, which are being procured, which are built in-house, which are embedded in other products.

Practical: a table with name, vendor, use case, data sources, user groups, business owner. Important: inventory Microsoft Copilot, ChatGPT Enterprise, Salesforce Einstein and all SaaS-embedded AI features.

Sources: SaaS inventory from SSPM, license analysis, business unit surveys, discovery via CASB tooling.

Step 2: Check Art. 5 prohibitions

Since 2 February 2025 Art. 5 prohibitions apply: social scoring by public or private actors, manipulative or exploitative AI, real-time remote biometric identification in publicly accessible spaces (with narrow law-enforcement exceptions), emotion recognition in workplaces and education, untargeted scraping of facial images.

Breach exposes up to 35 M EUR or 7% of annual turnover. In practice: HR AI with emotion recognition, employee monitoring with voice analysis, candidate screening with stress detection. Systems like these are prohibited since February 2025 and must be discontinued immediately.

Step 3: Check Annex III high-risk

Annex III lists the high-risk domains: critical infrastructure, education and training, employment (selection, evaluation, dismissal), access to essential services (credit scoring, insurance risk), law enforcement, migration / border control, justice, democratic processes, biometric categorisation.

Pragmatic test: is the AI used in one of these areas? Does it make substantial decisions about people? If yes, usually high-risk, to be confirmed against Art. 6 in detail.

Note: providers (system makers) and deployers (users) have different obligations. Classification is at the system level; obligations differentiate by role.

Step 4: Check Annex I high-risk

Annex I lists regulated products in which AI acts as a safety component: machinery, toys, lifts, pressure equipment, medical devices, in-vitro diagnostics, personal protective equipment, marine equipment, rail components, motor vehicles, civil aviation.

Application date here is 2 August 2027 (one year later than Annex III). Assessment usually runs in the product conformity-assessment process (CE marking).

Step 5: Check Art. 50 transparency

Even if not high-risk: transparency duties apply to chatbots (AI interaction notice), deepfakes (labelling), emotion recognition (informing affected persons), biometric categorisation, AI-generated content.

Practical: every generative AI workflow (marketing content, images, video, code) needs a labelling review. Which outputs go outside? Are they to be labelled? Who is responsible?

Documentation, documentation, documentation

The classification decision per system must be documented. Mandatory content: system name, provider, use case, assessed risk class with reasoning, derived obligations, owner, date. In a future supervisory review this documentation is your first evidence. DecisionOS structures exactly these classification decisions per AI system as a compact decision memo.

EU AI Act in practice: classify AI systems in five steps