The classic trap
Annex III is the list that reclassifies your AI project. Many organisations deploy a CV screening tool, a credit score, an exam proctoring system or an emotion recognition camera thinking they are in a banal use case. As soon as a case falls within one of the 8 Annex III domains, the system flips to high-risk and triggers the entire Chapter III (risk management, data quality, technical documentation, logging, transparency, human oversight, robustness, registration in the EU database). The EU AI Office and the CNPD for the personal data side will look at this qualification first, before everything else.
The Annex III qualification grid: the 8 domains to audit internally
- Biometrics: remote identification, categorisation on sensitive attributes, emotion recognition (1:1 identity verification is NOT covered).
- Critical infrastructure: safety components for water, gas, electricity, heating, road traffic, critical digital infrastructure.
- Education and training: access, admission, learning outcomes evaluation, orientation, exam monitoring.
- Employment and HR: recruitment, application filtering, promotion or termination decisions, task allocation, performance evaluation.
- Essential services: eligibility for social benefits, credit scoring (excluding fraud detection), life and health insurance pricing, emergency call dispatching.
- Law enforcement: victim risk assessment, polygraphs, evidence reliability, re-offending risk, profiling under Directive 2016/680.
- Migration, asylum, borders: migration risk assessment, asylum and visa application review.
- Justice and democratic processes: judicial decision support, election influence.
Concrete qualification pitfalls
Three frequent mistakes: thinking that an HR chatbot is neutral when it filters applications (Annex III.4.a); believing that a banking anti-fraud tool escapes Annex III when it is also used to deny credit (flips to III.5.b); deploying a university proctoring SaaS tool without realising it is exam surveillance (III.3.d). The Article 6(3) derogation (preparatory task, minor improvement, no influence on the final decision) is narrow and must be documented before placing on the market, not after an inspection.
How Luxgap automates this risk
Our Luxgap AI Risk Classifier eliminates the Annex III qualification grey zone in a few hours, where traditional law firms invoice weeks of legal analysis. The tool combines a specialised LLM agent trained on the entire AI Act, EU AI Office Q&A and EDPB guidelines, with native connectors to your IT mapping (Azure AI Foundry, AWS Bedrock, Vertex AI, Hugging Face, Odoo and ServiceNow registers) to automatically retrieve your real AI systems rather than the declared ones.
- Automatically scans your cloud and SaaS environments to detect AI models in production (Azure OpenAI, Bedrock, Vertex, Mistral, Anthropic, in-house models) and reconstructs their actual business use case via API call logs.
- Classifies each system according to the 4 AI Act levels (prohibited, high-risk Annex III, transparency, minimal risk) and justifies each qualification with article and recital citations.
- Automatically tests the Article 6(3) derogation by asking the 4 key questions (preparatory task, improvement of human result, pattern detection without decisional influence, profiling exception) and generates the enforceable non-applicability documentation.
- Detects silent flips: an internal chatbot that becomes an HR tool, an anti-fraud detector that becomes credit scoring, an analytical dashboard that ends up as an automated decision tool.
- Generates the Annex IV compliance file (technical documentation) pre-filled for each system qualified as high-risk, with a link to the Article 71 EU database.
- Produces a timestamped, cryptographically sealed PDF report, enforceable before the EU AI Office and the CNPD during an inspection.
Available as a complement to a Luxgap DPO or CISO mandate or as a dedicated SaaS module depending on your scope. Request a tailored quote and our teams will prepare a demonstration on your actual AI systems, with a free 48h blank audit to qualify your Annex III exposure before any engagement.
