Regulation (EU) 2024/1689 · Artificial Intelligence Act ● Annex III: — days

The Annex III High-Risk AI Systems List: All Eight Use Cases Explained with Examples

Annex III is eight pages that decide your entire compliance budget. Here is every category with the mid-market examples that actually appear in audits — and the Article 6(3) test that can move a system out of scope.

Renomira Team · 11 min read

Two lists decide whether the AI Act’s heaviest regime touches you: Article 5’s prohibitions (nothing survives those) and Annex III — the catalogue of high-risk use cases. Everything downstream of a high-risk classification is expensive: technical documentation, conformity assessment, human oversight, logging, registration, fundamental-rights impact assessments for certain deployers.

Professionals at a job fair — hiring is a flagship Annex III high-risk use case
Photo: News Oresund / CC BY 2.0 via Wikimedia Commons

Annex III organises into eight areas. Below, each with the plain-English trigger and the examples that show up in real mid-market inventories — not the textbook ones.

The eight Annex III areas

# Area The trigger, condensed Mid-market examples
1 Biometrics Remote biometric identification, categorisation of natural persons, emotion recognition from biometric data Face-match building access; emotion analysis in a video tool
2 Critical infrastructure AI as a safety component in managing critical digital infrastructure, road traffic, or water/gas/heat/electricity Anomaly detection in a regional utility’s grid control
3 Education and vocational training Access admission, evaluation of learning outcomes, exam proctoring AI admission screening at a private academy; automated grading
4 Employment and worker management Recruitment, application screening, decisions on promotion/termination/terms, task allocation, monitoring and evaluation CV parsers, interview scoring, productivity scoring, AI rostering
5 Essential private and public services Public-benefit eligibility, creditworthiness, insurance pricing/risk, emergency triage, migration/border Automated credit scoring for a fintech; claims triage
6 Law enforcement Risk of offending, polygraphs, evidence reliability assessment, criminal profiling Mostly public-sector; vendor tools sold into police
7 Migration, asylum and border control Risk assessments, authenticity verification of applications, polygraph-style analysis Document-authenticity AI in an immigration authority
8 Administration of justice and democratic processes Assisting judicial decisions on facts or law; election-finding influencing Legal-research AI pitched as “draft your defence”

The four that catch mid-market companies

Most mid-market inventories that turn up high-risk land in three of these: employment, essential services, and education. Biometrics appears whenever building access goes facial; infrastructure appears in utilities and logistics ops.

Concretely, these earn an Annex III 4 classification at a 250-person company:

  • A CV parser that ranks candidates or rejects on threshold — screening applicants is Annex III 4, whether it recommends or decides.
  • An AI that proposes whom to fire, whom to promote, or how to allocate shifts with real effect on terms.
  • Meeting-recording tools that score employee “engagement” or performance — monitoring and evaluation of workers, and if the scoring infers emotion, a prohibited-practice problem under Article 5(1)(f) before you even reach here.

And these sit in Annex III 5:

  • Any creditworthiness evaluation of natural persons — an e-commerce financing feature or lending marketplace, not just banks.
  • Insurance risk assessment and pricing for life and health cover.
  • Triage that allocates emergency-service resources.

The escape hatch: Article 6(3) — and its four walls

A system that matches an Annex III area is not automatically high-risk. Article 6(3) derails the classification where the system:

  1. performs a narrow procedural task;
  2. improves the result of a previous human activity;
  3. detects decision patterns or deviations from patterns without replacing or influencing the human assessment;
  4. performs a preparatory task to an assessment — or does pure pattern recognition on input data.

Two walls stop the escape. First, if the system profiles natural persons using natural-language processing to infer personal traits — or does any of the derogation-gated evaluation of persons — the derogation is unavailable: classification stays high-risk where Annex III applies to that person-evaluation context. Second, member states may not apply the derogation at all for systems used in education, employment, essential services and their appeal procedures where the system influences outcomes on persons. In practice: the derogation helps for back-office pattern-detection, not for CV ranking.

A useful reading: does this system look at a person and change what happens to them? If yes, assume high-risk and document the reasoning. Does it scan documents for anomalies that a human then judges? Argue 6(3)(c) — and write down why.

What high-risk actually costs

Scope, in effort rather than adjectives, for a deployer under Article 26:

  • Input data and instructions — assign human oversight, ensure competent operators.
  • Monitoring and logging — automatic logging of operation; Article 26(6) six-month minimum retention (longer where sector law demands it).
  • Retransmission discipline — check input data relevance per the provider’s instructions.
  • FRIA under Article 27 — if you are a public body, a private entity providing public services, or you run creditworthiness or life/health insurance pricing systems.
  • Registration under Article 49 — unless you are an SME (the registration step, not the duties, is what exempts).
  • Worker consultation — Article 26(7): inform workers’ representatives before workplace deployment.

A provider’s list is longer: Annex IV technical documentation, conformity assessment, CE marking, post-market monitoring, Article 73 incident reporting.

The classification record that survives review

Whatever the outcome, keep one artefact per system: what it does, which Annex III areas were considered, which Article 6(3) prongs were argued, what concluded, who signed, when. Three sentences and a date beat a 30-page framework — because the inspector reads the reasoning, not the label.

How to check your own position in four minutes

The free exposure scan asks twelve questions about your AI usage and governance and returns an exposure score, your ranked exposures with the exact article that fired, and your first three moves — mapped to the dates that apply. Confirm your email and the full report unlocks.

Frequently asked questions

What are the eight Annex III high-risk AI categories?

Biometrics; critical infrastructure; education and vocational training; employment and worker management; essential private and public services; law enforcement; migration, asylum and border control; and administration of justice and democratic processes. Each area has sub-points that define exactly which uses are high-risk — the sub-point text, not the heading, is what you classify against.

Is an AI that screens job applicants high-risk?

Yes — screening applicants falls under Annex III area 4 (employment), whether the system rejects, ranks or merely recommends. The Article 6(3) derogation is not available for systems influencing employment decisions about persons. Obligations include the provider’s conformity route and, on your side as deployer, oversight assignment, logging, worker consultation under Article 26(7) and registration unless you are an SME.

Can a system leave the high-risk category under Article 6(3)?

Sometimes. Narrow procedural tasks, human-improvement assists, and pattern-detection that neither replaces nor influences a human assessment can qualify. Two hard limits: profiling people to infer traits blocks the derogation, and member states may not allow it where the system influences decisions in education, employment, essential services or their appeal procedures. The classification reasoning must document whichever way it goes.

What is the difference between Article 5 prohibited and Annex III high-risk?

Article 5 bans practices outright — social scoring, untargeted facial-image scraping, workplace emotion inference — with no compliance route since 2 February 2025. Annex III high-risk systems are permitted but burdened: documentation, oversight, logging, conformity assessment and registration, with deployer duties from 2 December 2027. One is a stop sign; the other is a toll booth.

Do high-risk obligations apply to systems we bought, not built?

Yes. As a deployer, Article 26 obligations attach to your use: oversight, input-data checks, log retention, worker consultation, and FRIA where you fall in the Article 27 trigger set. You rely on the provider for technical documentation — which is exactly why procurement must capture their classification claim before signature.

Next step

Put this guide to work on your own estate.

Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.

  • Free forever plan
  • No credit card
  • EU-hosted · GDPR-native
Run a free scan Sign up free

Build a live AI register in minutes — upgrade only when you need the audit pack.

Annex III high-risk AI systems listEU AI Act high risk use cases examplesAI Act Annex III categorieswhich AI systems are high risk EUhigh-risk AI examples employment credit educationAnnex III derogation Article 6(3)AI Act high risk definition

Keep reading