Which AI systems are high-risk under Annex III?
Annex III describes eight areas containing specified AI uses. Check intended purpose and conditions before classifying a tool. A department or product name is insufficient.
What are the eight areas?
The list covers specified uses, not every piece of software in each sector:
- Biometrics: specified remote identification, categorisation and emotion recognition, where the use is permitted. Verification solely to confirm a claimed identity is distinguished from remote identification.
- Critical infrastructure: safety components in managing critical digital infrastructure, road traffic or supplies of water, gas, heating and electricity.
- Education and vocational training: specified decisions on access, assignment and educational level, evaluating outcomes and detecting prohibited behaviour during tests.
- Employment: specified recruitment, employment decisions, task allocation based on personal characteristics or behaviour, and performance assessment.
- Essential private and public services: specified public assistance, creditworthiness with an exception for financial fraud detection, life and health insurance, emergency calls and triage.
- Law enforcement: specified uses by or on behalf of competent authorities concerning victims, evidence and risk assessment, where permitted by applicable law.
- Migration, asylum and borders: specified assessments, applications and identification by or on behalf of competent authorities, subject to the legal limits.
- Justice and democratic processes: specified assistance to justice and influence on elections or referendums. Purely administrative campaign organisation without direct exposure of people to output is excluded from the relevant electoral use.
This points you to the detailed text; it is not a complete classification. A high-risk category does not permit a prohibited practice. Article 6 also has a separate route for Annex I products; the eight areas do not exhaust high-risk systems.
When is an Article 6 exception possible?
For an Annex III system, Article 6 allows an exception where there is no significant risk of harm to health, safety or fundamental rights, including through no material influence on the decision, and the relevant condition is met:
- A narrow procedural task.
- Improving the result of an already completed human activity.
- Detecting decision patterns or deviations without replacing or influencing the completed human assessment without proper human review.
- A preparatory task for assessment in the specified uses.
A person’s presence in the process alone does not establish the exception. An Annex III system profiling natural persons remains high-risk. That does not classify every instance of profiling by any AI as an Annex III system.
A provider relying on the exception documents its assessment before placing on the market or putting into service and has an Article 49 registration duty. A competent authority may request the documentation. A ready-made tool’s user does not automatically acquire the same duty merely by being a deployer.
Who assesses conformity?
For biometrics in point 1, Article 43 links the choice between internal control and a notified-body procedure to applied harmonised standards or common specifications. Missing, partial or restricted coverage can require the body procedure. Check the particular conditions.
Points 2–8 have internal control under Annex VI. Where the system is also covered by sectoral legislation in Annex I Section A, Article 43 provides for the relevant sectoral procedure. Do not treat internal control as universally sufficient for every product.
Check current standards and the notified body’s scope in the European Commission’s NANDO register of notified bodies. System registration is not a certificate of conformity.
When should you prepare classification?
Article 113 sets 2 December 2027 for the relevant Annex III high-risk rules and 2 August 2028 for the Annex I product route. These are not deadlines postponing the entire AI Act.
For systems already placed on the market or put into service, check the Article 111 transitions and significant design changes. Do not assume an older tool is unconditionally exempt or that one date settles every duty.
Organise intended purpose, role, category and the grounds for assessment while you can still request missing answers from the provider. Separate established information from disputed legal classification.
Questions
Is every AI tool in HR high-risk?
No. Check the particular use within the employment area, conditions and possible exception. Profiling by an Annex III system changes that assessment.
Do we always need a notified body?
No. The procedure depends on category, standards, specifications and possible sectoral overlap. Check the applicable route and current body scope.
We consider the system non-high-risk. What next?
Clarify your role. A provider relying on the exception documents the assessment and fulfils applicable registration duties. As a deployer, request the grounds and check your own duties; a tick in a list is not legal confirmation.
Sources
- European Commission — Annex III.
- European Commission — Article 6: classification.
- European Commission — Article 43: conformity assessment.
- European Commission — Article 113, as amended by Regulation (EU) 2026/1744: application dates.
- European Commission — Article 111: transitions.
- Regulation (EU) 2026/1744: amendments.
- European Commission: NANDO, register of notified bodies.
Related reading
Next step
Record the category and its grounds.
EU AI Act Readiness organises tools, uses and the category assessment with sources and dates. Disputed classifications are identified for legal review; the register and duty map do not certify compliance.
or write to contact@ergostrat.com