The classic trap
Recital 52 clarifies Annex III: a stand-alone AI system (not a safety component of a product) becomes high-risk as soon as it operates in one of the listed areas AND poses a high risk to health, safety or fundamental rights. The classic trap: believing that an HR tool, a credit scoring engine or a medical triage assistant is just a productivity SaaS, when it actually falls under Annex III. The EU AI Office and, for personal data aspects, the Luxembourg CNPD, expect a documented and reasoned qualification, not a verbal self-declaration.
The qualification methodology every provider or deployer must trace
Recital 52 implicitly imposes a reproducible classification methodology, since the Commission will use the same criteria to update Annex III via delegated acts. Concretely, for each AI system in your estate:
- Identify the intended purpose as documented in the notice and marketing, not the theoretical usage.
- Check whether the system falls into one of the eight Annex III areas (biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, justice and democracy).
- Assess the severity of potential harm (reversible / irreversible, individual / collective, systemic discrimination).
- Assess the probability of occurrence based on technical data (error rates, measured biases, deployment scope).
- Conclude with a reasoned qualification: high-risk, limited risk (art. 50) or minimal risk.
- Reassess at each substantial model update or change of use, because Article 6(3) provides exemptions from high-risk that may disappear.
The 'intended purpose' versus 'real use' test: the key argument
Recital 52 reasons on the intended purpose, not on the marketing pitch. The same generalist LLM can be high-risk at an HR deployer (CV screening) and out of scope at a marketing deployer (slogan generation). Qualification is done per use case, not per technology. This distinction is decisive when negotiating with your providers: require an explicit intended purpose sheet from them, reject vague 'general purpose' clauses.
How Luxgap automates this risk
Our Luxgap AI Risk Classifier turns AI Act qualification from an abstract legal exercise into a reasoned, traceable decision enforceable against the EU AI Office. The tool scans your IT estate (Azure AI Foundry, AWS Bedrock, Google Vertex, OpenAI Enterprise, Hugging Face Spaces, M365 Copilot, Salesforce Einstein) to inventory your real AI systems, then automatically applies the Recital 52 methodology to propose a reasoned classification article by article.
- Automatically detects each AI deployment in your Microsoft 365 tenant, your Azure / AWS / GCP workloads and your connected SaaS applications, without manual declaration from business units.
- Cross-references the declared intended purpose (extracted from vendor contracts and technical sheets) with the eight Annex III areas to determine high-risk classification.
- Computes a documented severity x probability score, based on model error rate, impacted population and harm reversibility.
- Generates a timestamped, cryptographically sealed PDF qualification sheet, ready to produce to the EU AI Office or the CNPD during an audit.
- Sends real-time Teams or Slack alerts as soon as a new model is deployed or a substantial update changes the risk profile.
- Maintains automated monitoring of Commission delegated acts amending Annex III and reclassifies your estate accordingly.
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 real AI estate, with a free 48-hour blind audit to map your high-risk systems before any commitment.
