The Filter and the File: How the EU Just Redrew High-Risk AI Classification

What happened. The Commission published three texts under Article 6(5): general principles, the Annex I product-safety route, and the Annex III use-case route, plus a list of practical examples. The Annex I text gives the methodology a few hard edges. A failed thermostat means inconvenience and a higher bill, so most smart appliances stay out, while a failed smart door lock or child-lock can endanger a child, so those cross into high-risk. The consultation runs until June 23 (22:00 CET) through an anonymous questionnaire, and the AI Act Single Information Platform hosts a searchable explorer of the examples.

Why it matters. Before May 19, companies were classifying against statutory language with no Commission reading behind it, which left the most consequential question under the Act unsettled, stalled internal sign-offs, and pushed cautious teams to over-classify everything in sight. The examples are explicitly non-exhaustive, and the Commission says that listing a use case settles nothing about whether it is lawful. The July 2025 GPAI guidance showed how this goes: a non-binding text becomes the reference that regulators, auditors, and counterparties reach for within weeks of publication.

What happened. Article 6(2) makes any system in one of eight Annex III areas high-risk: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, and the administration of justice. Article 6(3) lets a system escape if it meets one of four conditions, namely that it performs a narrow procedural task, improves the result of a completed human activity, detects decision patterns without replacing human judgment, or carries out a preparatory task. The Commission then bolts on two gates. The system must not materially influence the outcome, and it must not perform profiling. Because the filter is an exception to rules that protect fundamental rights, the guidance states that its conditions have to be read narrowly.

Why it matters. Bird & Bird calls 6(3) the single most asked-about provision in the whole Annex III analysis. The employment examples show where the line falls, and they reach into every sector that hires. The Commission treats systems that source, score, rank, shortlist, assess, or match candidates as high-risk, while tools that schedule interviews, verify credentials, or parse CVs into a searchable database may fall outside. Enzai’s illustration is the cleanest one to keep in mind. A CV tool that sorts applications into predefined categories can pass the procedural-task exemption, and the same tool tuned to score or rank suitability cannot, even where a human reads the result.

What happened. The escape is an active obligation. Article 6(4) requires the provider to document the assessment before the system reaches the market, and the system goes into the EU database under Article 71. Market surveillance authorities can call for that file whenever they choose, reclassify the system under Article 80, and reach for Article 99 penalties where a misclassification looks like an attempt to dodge the regime.

Why it matters. The asymmetries in the text reward close reading, because they decide real money. Credit scoring is high-risk, with a narrow carve-out for financial-fraud detection, while health and life insurance risk assessment and pricing carries no equivalent carve-out, a gap insurers running fraud and underwriting on shared infrastructure will feel right away. For modular and agentic builds, the Commission assesses the whole configuration where the components’ joint outputs materially influence a decision, so splitting functionality across modules to slip under the threshold gets nowhere.