Annex III of the AI Act lists eight use areas in which AI systems are classified as high-risk under Article 6(2). The Commission guidelines of 19 May 2026 provide a per-domain breakdown of which use cases qualify as high-risk, with concrete examples from practice. This overview summarises the core of each domain and links to the deeper analyses per area.
For the general structure of the Article 6(3) filter and the three pitfalls that apply across all domains, see the main article on how the Article 6(3) filter works.
For each domain you'll see: the relevant use cases under Annex III, examples of systems that qualify as high-risk, and examples that may fall within the filter. The lists are not exhaustive; the Commission indicates they will be updated over time. Treat this page as a starting point for your own classification, not a final verdict.
Test your AI against the Article 6(3) filter
Free interactive self-assessment, updated for the Commission guidelines of 19 May 2026. 9 steps, personal report with reasoning, vendor questions and next steps.
1. Biometrics
The biometrics domain covers three use cases: remote biometric identification, biometric categorisation and emotion recognition. It is one of the most detailed domains in the guidelines, with over twenty pages of interpretation and examples.
Key use cases:
- Point 1(a) Remote biometric identification: systems that identify persons at a distance based on biometric data, without their active cooperation
- Point 1(b) Biometric categorisation: systems that classify persons into groups based on sensitive biometric features
- Point 1(c) Emotion recognition: systems that detect emotions or mental states from biometric signals
The guidelines clarify that 1-to-1 verification (confirming a claimed identity) falls outside Point 1(a), while 1-to-many identification is within scope. For emotion recognition, the boundary between prohibited practices (Article 5) and high-risk (Annex III) is critical.
Read the deep-dive on high-risk AI in biometrics
2. Critical Infrastructure
This domain covers six use cases for AI systems functioning as safety components in essential services. The Commission guidelines connect this closely with the NIS2 Directive and the CER Directive on critical entities.
Key use cases:
- Safety components in critical digital infrastructure
- Road traffic (traffic lights, intelligent transport systems)
- Water supply
- Gas supply
- Heating supply
- Electricity supply
The guidelines emphasise that only systems functioning as safety components fall within scope. AI used purely for supporting functions (administration, planning without safety function) falls outside high-risk classification.
Read the deep-dive on high-risk AI in critical infrastructure
3. Education and Vocational Training
This domain has four use cases that touch all stages of the education process: admission, evaluation, level determination and behaviour detection.
Key use cases:
- Point 3(a) Determining access or admission to educational institutions
- Point 3(b) Evaluating learning outcomes and steering the learning process
- Point 3(c) Determining the appropriate level of education
- Point 3(d) Monitoring and detecting prohibited behaviour during tests
The guidelines distinguish between pedagogical support (often filter-eligible) and evaluative applications that directly affect a student's future (always high-risk). For proctoring systems, the boundary lies in whether the system itself detects behaviour or merely supports the teacher.
Read the deep-dive on high-risk AI in education
4. Employment and Worker Management
The employment domain has two broad use cases that together cover the entire employee lifecycle, from recruitment to contract termination.
Key use cases:
- Point 4(a) Recruitment and selection of natural persons (job description to final decision)
- Point 4(b) Management of work-related relationships (task allocation, evaluation, promotion, termination)
This is one of the most commercially relevant domains. Virtually any HR-tech system that ranks, scores or selects candidates falls within high-risk. The guidelines go into detail on the boundary between sourcing tools (often out of scope or filter-eligible) and screening/ranking tools (always high-risk).
For the broader impact on the HR sector, see also our earlier analysis on the silent revolution in HR and recruitment.
Read the deep-dive on high-risk AI in employment
5. Essential Services
This domain bundles four very different use cases, with the common thread that they govern access to services that profoundly affect daily life.
Key use cases:
- Point 5(a) Eligibility assessment for public benefits and services
- Point 5(b) Creditworthiness and credit scoring
- Point 5(c) Risk assessment and pricing in life and health insurance
- Point 5(d) Triage and prioritisation of emergency calls and dispatch
For the financial sector, the guidelines provide important clarification on the interaction with Article 144 CRR (Capital Requirements Regulation) and Article 120 Solvency II. Banks and insurers should look at both carefully.
For the financial sector specifically, see our article on AI governance in the financial sector.
Read the deep-dive on high-risk AI in essential services
6. Law Enforcement
The law enforcement domain has five use cases plus clear delineation of what falls outside scope. The guidelines emphasise that many AI applications in policing fall under prohibited practices (Article 5), and that high-risk classification becomes relevant only when the application falls outside that.
Key use cases:
- Point 6(a) Assessing the risk of a person becoming the victim of a crime
- Point 6(b) Polygraphs and similar tools
- Point 6(c) Assessing the reliability of evidence
- Point 6(d) Assessing the risk of (re)offending by concrete persons
- Point 6(e) Profiling of natural persons during criminal investigation
Political and operational delineation is critical here: forensic analysis of a specific crime scene often falls outside scope, but predictive systems targeting persons or neighbourhoods usually sit in high-risk or prohibited zone.
Read the deep-dive on high-risk AI in law enforcement
7. Migration, Asylum and Border Control
This domain has four use cases covering the entire chain of admission to Europe. The guidelines connect it closely to existing Schengen, EES, ETIAS and EURODAC systems.
Key use cases:
- Point 7(a) Polygraphs and similar tools in migration context
- Point 7(b) Risk assessment of persons seeking to enter or stay in the EU
- Point 7(c) Assistance in examining asylum, visa or residence applications
- Point 7(d) Detection, recognition or identification of persons in migration or border context (excluding travel-document verification)
The boundary between administrative support (filter-eligible) and substantive assessment (always high-risk) is the central interpretive question here.
Read the deep-dive on high-risk AI in migration and asylum
8. Administration of Justice and Democratic Processes
The eighth and final domain consists of two very different use cases that touch the pillars of democratic rule of law.
Key use cases:
- Point 8(a) AI systems supporting judicial authorities or alternative dispute resolution
- Point 8(b) AI systems intended to influence the outcome of elections or referendums
For Point 8(a), the guidelines are strict: any system that substantively contributes to investigation or interpretation of facts by a judge is high-risk. Pure administrative assistance (calendar, document management) falls outside scope.
For Point 8(b), the guidelines are nuanced: legitimate campaign tools and political communication are not included, but systems specifically intended to influence election outcomes, for example via microtargeting or synthetic media, are within scope.
Read the deep-dive on high-risk AI in justice and democracy
What This Overview Gets You
For each of the eight domains the same three-step analysis applies:
Does the use case fall within Annex III?
Start from actual use. Which decisions or assessments does the system support, and who is impacted? Intended purpose is decisive, not the technology.
Does the Article 6(3) filter apply?
Per use case within Annex III: can the system fall within one of the four filter conditions (narrow procedural, ex-post improvement, pattern detection, preparatory)? Or is there profiling (which automatically excludes the filter)?
Document and register
Whether the system is high-risk or falls outside via the filter, both conclusions must be documented. Filter status must be registered in the EU database.
For details per domain, use the links in each section above. For the general framework, the filter and the three pitfalls, see the main article on the Article 6(3) filter.
Practical Order for Your AI Portfolio
Which domains to scrutinise first depends on your sector. A few rules of thumb:
- Employers and HR-tech: start with domain 4 (employment)
- Financial institutions: start with domain 5 (essential services), with specific attention to the interaction with CRR and Solvency II
- Insurers: domain 5(c) is your focus, plus possibly domain 1 if you use biometrics for fraud detection
- Educational institutions and EdTech: domain 3 (education)
- Public sector (municipalities, agencies): domains 5(a), 6 (law enforcement) and 8(a) (justice)
- Energy, water, telecom: domain 2 (critical infrastructure)
- Border, immigration, customs authorities: domain 7 (migration and border control)
- Online platforms and media companies: domain 8(b) (democratic processes)
For a structured approach to AI inventory and risk classification, our decision tree for risk classification provides a first interactive check. For team-wide AI literacy under Article 4, which forms the basis for consistent classifications, LearnWize is the logical next step.
Test your AI against the Article 6(3) filter
Free interactive self-assessment, updated for the Commission guidelines of 19 May 2026. 9 steps, personal report with reasoning, vendor questions and next steps.
Frequently Asked Questions
Questions about the structure and application of Annex III high-risk classification.