Skip to content
SimpleAct Logo

Annex III AI Act · Article 6

Updated 7 October 2026 · after the Digital Omnibus

EU AI Act Annex III: all high-risk areas with examples

Annex III lists eight areas in which AI systems count as high-risk – from biometrics to recruitment to credit scoring. Here you will find each of the 25 use cases with a typical example and a counter-example, a check for the Article 6(3) exemption and a documentation template to copy.

379 days until the Annex III obligations apply (2 Dec 2027)

All 25 use cases with an example · Counter-example: when it is not high-risk · Article 6(3) check · Obligations and deadlines after the Omnibus

app.simpleact.de

Risk classification

AI inventory · HR

A
Applicant screeningScores and ranks incoming applications
  • Annex III areaPoint 4 – Employment
  • Profiling of natural personsyes
  • Article 6(3) exemptionnot applicable

Result

High-riskAnnex III point 4(a)

Resulting duties

Art. 26 deployerArt. 26(7) inform staffDPIA Art. 35 GDPRfrom 2 Dec 2027

Assessment documented · 3 Oct 2026

In short

Annex III of the EU AI Act names eight areas with 25 specific use cases: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, and justice and elections. A system used there is high-risk unless it only performs an ancillary task under Article 6(3) and does no profiling. The obligations apply from 2 December 2027.

Context

Two routes to high-risk AI

Article 6 of the AI Act has two independent routes. Annex III is the route via the intended purpose – it affects most companies using AI in HR, finance or customer contact. Annex I is the route via regulated products.

Two routes to high-risk AI
FeatureRoute 1: Annex I (Art. 6(1))Route 2: Annex III (Art. 6(2))
Linked toThe product the AI is part ofThe purpose the AI is used for
When high-riskThe AI is a safety component of a product, or is itself a product, covered by the harmonisation legislation in Annex I – and the product needs third-party assessmentThe intended purpose falls under one of the 25 use cases in the eight areas of Annex III
Typical examplesAI in medical devices, machinery, lifts, toys or radio equipmentApplicant screening, credit scoring, online proctoring, emergency call prioritisation
Exemption possibleNo – Article 6(3) only applies to Annex IIIYes, under Article 6(3) – but never with profiling of natural persons
Conformity assessmentWithin the procedure under the relevant product legislationFor points 2 to 8 usually internal control by the provider; for biometrics (point 1) possibly involving a notified body
Obligations apply from2 August 2028 (originally 2 August 2027)2 December 2027 (originally 2 August 2026)

Simplified overview. A system can be caught by both routes; the requirements then apply once, but checked from both angles.

The full overview

All 8 areas and 25 use cases in Annex III

For each point: what is covered, a typical real-world example, and where a similar system is probably not high-risk. Filter by area or expand and collapse areas.

All 8 areas and 25 use cases in Annex III

Biometric systems are only covered in so far as their use is permitted under Union or national law at all. Some biometric practices are already prohibited under Article 5 – such as emotion recognition at work and in education (except for medical or safety reasons) or categorisation that infers political opinions, religion or sexual orientation.

1(a)Remote biometric identification: people are matched against a reference database at a distance and without their active involvement.Typical exampleCamera system that matches visitors to a stadium or shopping centre against a list of banned persons.Not high-risk if …the system only confirms that someone is who they claim to be – such as Face ID on a work phone or matching a face against one’s own ID card at an access gate (pure verification).
1(b)Biometric categorisation: people are sorted based on biometric data according to inferred sensitive or protected attributes.Typical exampleIn-store video analytics that groups customers by estimated age and sex from facial images.Not high-risk if …the grouping is not linked to sensitive or protected attributes, or is merely a technically necessary ancillary feature of another service, such as a photo filter that detects facial features.
1(c)Emotion recognition: the system infers emotions or intentions from face, voice or body language.Typical exampleAnalysis of voice and tone in call-centre conversations to detect the mood of callers.Not high-risk if …only obvious expressions such as a smile are detected without inferring feelings, or the system concerns physical states such as fatigue. Note: emotion recognition at work and in education is prohibited in principle.

What matters here is the role as a safety component: a component whose failure or malfunction can physically endanger the infrastructure or threaten people’s health and safety. According to the recitals, components used solely for cybersecurity are not safety components in this sense.

2Safety components in the management and operation of critical digital infrastructure, road traffic, and the supply of water, gas, heating and electricity.Typical exampleAI that controls load flows in a distribution grid and switches off grid sections on its own when overload looms; AI-controlled traffic lights.Not high-risk if …the system is not a safety component, so an error endangers neither people nor supply – such as AI for billing, customer service or analysing consumption data for reports.

Educational institutions at all levels are covered, from schools to universities to vocational training. The question is whether the AI co-determines access, assessment or placement of learners – because that shapes educational and career paths.

3(a)Determining access, admission or assignment to educational institutions.Typical exampleSoftware that pre-sorts university applications and produces a ranking for admission.Not high-risk if …the system only checks documents for completeness or schedules selection interviews – a narrow procedural task.
3(b)Evaluating learning outcomes, including where this is used to steer the learning process.Typical exampleAutomated marking of exams; a learning platform that uses test results to decide which modules someone must take next.Not high-risk if …the AI only polishes the wording of feedback the teacher has already written, or provides practice exercises without assessment.
3(c)Assessing the appropriate level of education a person will receive or be able to access.Typical examplePlacement test that automatically decides the level of a language course or access to a support programme.Not high-risk if …the result remains a non-binding self-assessment with no effect on access or placement.
3(d)Monitoring and detecting prohibited behaviour of students during tests.Typical exampleOnline proctoring that analyses eye movements via webcam and flags possible cheating.Not high-risk if …the system only secures the exam environment technically (such as a browser lock) and does not assess candidates’ behaviour.

The most relevant area for companies in practice. It runs from recruitment to dismissal and includes platform work. Many HR tools now contain AI features – what matters is whether they assess, select or steer people, or merely ease administrative work.

4(a)Recruitment and selection: placing targeted job ads, analysing and filtering applications, evaluating candidates.Typical exampleApplicant tracking software that scores CVs for fit and produces a ranking for recruiters; video interview analysis.Not high-risk if …the AI only drafts job descriptions from the team’s input – a narrow procedural task according to the draft guidelines – or coordinates interview slots.
4(b)Decisions on terms of work, promotion and termination, allocating tasks based on behaviour or personal traits, monitoring and evaluating performance and behaviour.Typical exampleSystem that allocates jobs to warehouse or delivery staff based on individual performance and analyses metrics for appraisals.Not high-risk if …shifts are planned solely by availability and formal qualifications without assessing behaviour or personality, or the AI only translates employment contracts.

This area concerns services people rely on to lead an independent life: public benefits, credit, life and health insurance and emergency help. For banks, insurers and fintechs it is central.

5(a)Public authorities evaluate eligibility for public assistance benefits and services, including healthcare, and grant, reduce, revoke or reclaim them.Typical exampleSystem at a social welfare authority that assesses benefit applications and proposes reductions or reclaims.Not high-risk if …the AI only routes incoming mail to the responsible unit or checks applications for missing information.
5(b)Evaluating the creditworthiness of natural persons or establishing their credit score – except systems for detecting financial fraud.Typical exampleScoring model of a bank or buy-now-pay-later provider that decides on credit limits for consumers.Not high-risk if …the system only detects fraud patterns in transactions (expressly excluded) or rates the creditworthiness of companies rather than natural persons.
5(c)Risk assessment and pricing for natural persons in life and health insurance.Typical examplePricing model that calculates private health insurance premiums individually from health and lifestyle data.Not high-risk if …it concerns motor, home contents or other property insurance, or the AI only sorts incoming claims.
5(d)Evaluating and classifying emergency calls, dispatching or prioritising first responders such as police, fire brigade and ambulance, and patient triage in emergency care.Typical exampleControl-room software that ranks emergency calls by urgency and proposes the order of ambulances; initial triage in an emergency department.Not high-risk if …the system only transcribes calls or shows locations on a map without influencing prioritisation.

Covered are systems used by or on behalf of law enforcement authorities – and only in so far as permitted under Union or national law. For companies this area matters mainly as providers of such solutions. Some practices, such as crime predictions based solely on profiling, are prohibited under Article 5.

6(a)Assessing the risk of a person becoming the victim of a criminal offence.Typical exampleAnalysis tool that estimates which persons are at particular risk of becoming victims of domestic violence.Not high-risk if …only anonymous situation reports without reference to individuals are produced.
6(b)Polygraphs and similar tools.Typical exampleSoftware meant to infer the credibility of statements from voice or facial expressions during interviews.Not high-risk if …hardly any leeway: such tools are aimed directly at assessing people.
6(c)Evaluating the reliability of evidence.Typical exampleAI that assesses the authenticity of photos, videos or documents in an investigation.Not high-risk if …the AI only indexes evidence files, makes them searchable or removes duplicates.
6(d)Assessing the risk of offending or re-offending, not solely based on profiling, or assessing personality traits and past criminal behaviour.Typical exampleRecidivism tool that proposes a risk level to probation officers based on objective facts.Not high-risk if …none of this applies. Note: predictions based solely on profiling or personality traits are prohibited under Article 5(1)(d).
6(e)Profiling of natural persons in the detection, investigation or prosecution of criminal offences.Typical exampleAnalysis of communication and movement data to filter out possible suspects.Not high-risk if …no exemption: under Article 6(3), profiling is always high-risk.

Covered are systems used by or on behalf of competent public authorities and Union bodies, in so far as legally permitted. The people affected are in a particularly vulnerable position – which is why the list is comparatively broad here.

7(a)Polygraphs and similar tools.Typical exampleBorder interview tool that automatically scores answers for credibility.Not high-risk if …hardly any leeway: such tools are aimed directly at assessing people.
7(b)Assessing risks allegedly posed by a person who intends to enter or has entered – such as security, irregular migration or health risks.Typical exampleSystem that automatically assigns travellers a risk level before entry.Not high-risk if …only aggregated travel data are analysed for staff planning at border crossings.
7(c)Assisting authorities in examining applications for asylum, visas and residence permits and related complaints, including assessing evidence.Typical exampleAI that pre-assesses visa applications and points out possible grounds for refusal.Not high-risk if …the AI only translates documents or routes applications to the responsible units – preparatory tasks that must be documented.
7(d)Detecting, recognising or identifying persons in the context of migration, asylum and border control – except verifying travel documents.Typical exampleCamera or drone analysis that detects and identifies people in border areas.Not high-risk if …only the authenticity of a travel document is checked – this is expressly excluded.

Courts and elections are at the core of the rule of law. Covered is AI that supports courts on substance or is meant to influence elections – not purely organisational help, such as anonymising judgments or the logistics of a campaign.

8(a)Assisting judicial authorities in researching and interpreting facts and the law and applying the law to a concrete case; likewise in alternative dispute resolution.Typical exampleSystem that proposes decisions to a court or arbitration body based on the case file.Not high-risk if …the AI performs purely administrative support, such as anonymising judgments, scheduling or internal communication.
8(b)Influencing the outcome of an election or referendum or people’s voting behaviour.Typical exampleSystem that targets voter groups with personalised political messages to change how they vote.Not high-risk if …campaigns are only supported organisationally or logistically and nobody is directly exposed to the output – such as route planning for campaign stands.

Paraphrased summary of Annex III, not the legal wording. The right-hand column shows typical boundaries and is no substitute for a case-by-case check – what counts is the intended purpose set by the provider, and the Article 6(3) exemption must be documented.

Article 6(3)

The exemption: when an Annex III system is not high-risk

Not every system that touches an Annex III area is automatically high-risk. Article 6(3) exempts systems that do not pose a significant risk to health, safety or fundamental rights – in particular because they do not materially influence the outcome of a decision.

The exemption applies if at least one of four conditions is met. They describe systems that only touch a decision about people at the margins: they do ancillary work, refine a finished human result, or prepare something a human then assesses.

The assessment is made by the provider, i.e. whoever develops the system and places it on the market under their own name. They must document it before placing it on the market (Art. 6(4)) and still register the system in the EU database. Deployers can rely on this classification but should check it for plausibility – and anyone who uses a system for a purpose other than the intended one may become a provider under Article 25.

Look at actual practice: a feature that only “suggests” on paper but is almost always followed in everyday use does influence the decision. The exemption should therefore be read narrowly and backed up with concrete review steps.

  1. (a)

    Narrow procedural task

    The system performs a clearly defined, technical step – such as structuring unstructured data, sorting documents by type or finding duplicates.

    Example: According to the draft guidelines: AI that generates job descriptions from human-provided input.

  2. (b)

    Improving a completed human activity

    The AI builds on a result a human has already finished and improves it without changing the substance of the assessment.

    Example: Language editing of an expert report or employment reference that has already been written.

  3. (c)

    Detecting patterns or deviations

    The system detects decision-making patterns or deviations from them, but is not meant to replace or influence a previously completed human assessment without proper human review.

    Example: A subsequent flag that a teacher’s grading deviates from their usual pattern – the teacher checks it themselves.

  4. (d)

    Preparatory task

    The AI prepares an assessment relevant to an Annex III use case but does not make it – such as indexing, searching, translation or file processing.

    Example: Translating application documents in another language before a recruiter reads and assesses them herself.

Counter-exception: profiling is always high-risk

If an Annex III system performs profiling of natural persons – i.e. automatically evaluates or predicts personal aspects such as work performance, economic situation, health, interests, reliability or behaviour (Art. 4(4) GDPR) – it is always high-risk. None of the four conditions helps then.

Interactive

Check: does the Article 6(3) exemption apply?

Answer up to four questions about a specific system. At the end you get a first assessment, what to document and what applies to registration.

Question 1
Does the purpose of the system fall under a use case in Annex III?

Use the table above. What counts is the intended purpose set by the provider, not the technology used.

app.simpleact.de

Assessment

Art. 6(3) AI Act

A

Does the purpose of the system fall under a use case in Annex III?

First orientation, not legal advice. The provider is responsible for the classification and must document it.

Template

Documentation template for the Article 6(3) exemption

Anyone who classifies an Annex III system as not high-risk must document the assessment before placing it on the market and provide it on request. This outline covers the points an authority is likely to ask about. Copy it, paste it into your document, fill it in.

An outline, not a legally binding form. The reasoning must describe the actual use – phrases like “only supportive” are not enough.

Assessment-Article-6-3.txt
Documentation of the assessment under Article 6(3) and (4) AI Act
(SimpleAct template – adapt to your case, not legal advice)

1. AI system
- Name and version:
- Provider (name, address):
- Intended purpose (word for word from the instructions for use):
- Date of assessment, person responsible:

2. Annex III use case concerned
- Point and letter (e.g. point 4(a)):
- Why the use case is concerned in principle:

3. Profiling
- Does the system evaluate or predict personal aspects of natural persons (profiling within the meaning of Article 4(4) GDPR)? yes / no
- Reasoning (if "yes", the exemption is excluded):

4. Condition relied on under Article 6(3)
[ ] (a) narrow procedural task
[ ] (b) improves the result of a previously completed human activity
[ ] (c) detects decision-making patterns or deviations without replacing or influencing the human assessment without proper human review
[ ] (d) preparatory task to an assessment
- How exactly the system meets the condition:

5. No significant risk to health, safety or fundamental rights
- Influence of the output on the outcome of the decision:
- Groups of persons affected and fundamental rights concerned:
- Role of the human, review steps before the decision:
- Consequences of a system error for the persons affected:

6. Limits and triggers for a new assessment
- Functions deliberately not included:
- Changes that trigger a new assessment (new function, new purpose, new data sources, new user group):

7. Registration and disclosure
- Registered in the EU database under Article 49(2) on / ID:
- Where this documentation is kept, to be provided to the competent authority on request (Article 6(4)):

Approval: name, role, date

How to make the documentation hold up

  1. 01Take the intended purpose word for word from the instructions for use – the classification depends on it.
  2. 02Describe the real workflow: who sees the output, what happens next, who decides?
  3. 03Pick only a condition that genuinely fits and back it with an example from operations.
  4. 04Check profiling explicitly and explain why there is none.
  5. 05Define triggers for reassessment and version the documentation.
  6. 06Add the date and ID of the EU database registration.

Obligations

High-risk AI obligations: providers and deployers

If a system is high-risk under Annex III, the AI Act allocates obligations by role. Most fall on the provider. Anyone who merely uses a purchased system is a deployer – with a separate, smaller but binding set of duties.

Scroll the table sideways →

High-risk AI obligations: providers and deployers
ObligationProviderDeployerArticle
Risk managementSet up, document and maintain a system across the whole life cycle–Art. 9
Data and data governanceTraining, validation and test data meeting quality criteriaInput data under own control must be relevant and sufficiently representative for the purposeArt. 10, 26(4)
Technical documentationDraw up under Annex IV before placing on the market and keep up to date–Art. 11
Record-keepingTechnically enable automatic loggingKeep logs under own control for at least six monthsArt. 12, 26(6)
Transparency and instructionsInstructions for use with purpose, performance limits and oversight measuresUse the system in accordance with the instructionsArt. 13, 26(1)
Human oversightDesign the system so that people can oversee it effectivelyAssign competent people with authority and supportArt. 14, 26(2)
Accuracy, robustness, cybersecurityAchieve and declare an appropriate level–Art. 15
Quality managementDocumented quality management system–Art. 17
Conformity assessment and CEBefore placing on the market; for points 2–8 internal control (Annex VI), for point 1 (biometrics) possibly with a notified body (Annex VII); then EU declaration of conformity and CE marking–Art. 43, 47, 48
RegistrationRegister the system in the EU database – also with the Art. 6(3) exemptionPublic authorities and Union bodies register their useArt. 49
Fundamental rights impact assessment–Public bodies, private entities providing public services, and deployers of systems under points 5(b) and (c) – before first useArt. 27
Informing staff and affected persons–Inform workers and their representatives before use at the workplace; inform affected persons when the system makes or assists decisions about themArt. 26(7) and (11)
Monitoring in operationPost-market monitoring system with a planMonitor operation, inform the provider of risks and suspend useArt. 72, 26(5)
Serious incidentsReport to the market surveillance authorityInform the provider and the authority without delayArt. 73, 26(5)

Simplified overview. Anyone who puts their name on a high-risk system, substantially modifies it or changes its purpose so that it becomes high-risk is considered a provider under Article 25.

In practice most companies are deployers: they buy an HR tool, a scoring solution or control-room software and use it under their own responsibility. Their duties are manageable but organisationally demanding – above all human oversight by trained people, informing staff and keeping logs.

Providers carry the burden of product development: risk management, data quality, technical documentation, quality management and conformity assessment. For software companies building AI features into HR, finance or education products, Annex III classification is therefore a product decision that should be made early.

€15m
or 3% of worldwide annual turnover – maximum for breaches of high-risk obligations (Art. 99(4))
SMEs
For SMEs, including start-ups, the lower of the two amounts applies (Art. 99(6)), and since the Digital Omnibus also for small mid-caps (Art. 99(6a))
6 months
Minimum log retention for deployers, where logs are under their control

Deadlines

When do the obligations apply? Deadlines after the Omnibus

The Digital Omnibus, Regulation (EU) 2026/1744, in force since 27 July 2026, postponed the high-risk obligations. Annex III gets more time, but that is no reason to wait: classification, documentation and contracts need lead time.

  1. 2 December 2027

    Annex III

    Obligations for Annex III high-risk AI apply.

    originally 2 August 2026

  2. 2 August 2028

    Annex I

    Obligations for AI in regulated products under Annex I apply.

    originally 2 August 2027

  3. Legacy systems

    Placed on the market before

    Obligations generally only kick in if the system is significantly changed afterwards.

  4. 2 August 2030

    Systems for public authorities

    High-risk AI intended for use by public authorities must comply by then.

The structure is unchanged: Article 6 together with Annex III determines what is high-risk; Article 6(3) allows the exemption. Registration in case of an exemption also remains mandatory – the Omnibus provides a simplified procedure for it.

The legacy rule sounds generous but has a catch: for continuously developed software, a significant change is reached quickly. Anyone introducing a system with an Annex III link today should build the requirements into tenders and contracts already.

Guidelines

Commission guidelines on classification (draft)

Article 6(5) requires the Commission to provide guidelines with practical examples on classifying high-risk AI – originally by 2 February 2026. As of 7 October 2026, only a draft dated 19 May 2026 is available; the Commission has not published a final version. The draft is not binding, but shows how the Commission intends to read Article 6.

  1. 19 May 2026

    Draft guidelines published

  2. 23 Jul 2026

    Targeted consultation closed (extended from 23 Jun)

  3. End of 2026

    Adoption of the final version planned by the Commission

  4. 2 Dec 2027

    Annex III obligations apply

The draft goes through the Annex III areas and maps examples to them. One example is particularly relevant for HR teams: AI that generates job descriptions based on human-provided input is classified as a narrow procedural task under Article 6(3) – so not high-risk, even though it works in recruitment.

That is no free pass for HR tools as a whole. As soon as the same tool scores, sorts or filters applications, it is back under Annex III point 4(a). The guidelines help draw the line but do not replace checking the specific functions.

Until the final version, we recommend basing classifications on the text of the Regulation, citing the draft as an interpretative aid, and reviewing the assessment once the final version is published.

Key points of the draft (paragraph numbers)

  • The guidelines are not binding. Only the Court of Justice of the EU can give an authoritative interpretation of the AI Act. (para. 6)
  • A broadly framed intended purpose that does not consistently exclude high-risk uses makes the system high-risk. A mere exclusion in the terms of service is not enough if marketing or examples promote such uses. (para. 12)
  • Human involvement does not change the classification under Article 6(2). Adding a human in the loop does not turn an Annex III system into a low-risk one. (paras. 70–71)
  • The Article 6(3) exemptions are exhaustive and must be interpreted narrowly; the provider decides on them in a self-assessment. A system that performs profiling always remains high-risk. (paras. 88–89, 113)
  • Several AI components that together materially influence an individual decision are assessed as one system – agentic systems included. Splitting a system into modules does not avoid the classification. (paras. 75, 90)

What you can do now

  • Flag AI systems with a link to an Annex III area in your AI register
  • Create a documented Article 6(3) assessment for borderline cases
  • Set a reminder for the final version of the guidelines
  • Ask providers about their own classification and registration
Draft guidelines (European Commission)

GDPR

High-risk AI and data protection: DPIA, profiling, automated decisions

Most Annex III systems process personal data. The AI Act does not replace the GDPR but applies alongside it. Linking the two well saves duplicate work.

Data protection impact assessment (Art. 35 GDPR)

Annex III systems often assess people systematically and at scale – a typical trigger for a DPIA. Deployers should use the information from the provider’s instructions for use for this (Art. 26(9) AI Act).

DPIA guide

Fundamental rights impact assessment (Art. 27 AI Act)

Where a fundamental rights impact assessment is required, it complements a DPIA already carried out. Both can be kept in one document if the additional fundamental rights aspects are properly covered.

Profiling: one term, two laws

Article 6(3) refers directly to the GDPR definition of profiling. If profiling is already recorded in your record of processing activities, that is an important pointer for classification.

Automated individual decisions (Art. 22 GDPR)

If a system decides without human involvement with legal or similarly significant effect – such as on a loan – the limits of Article 22 GDPR also apply. Human oversight under the AI Act helps but does not replace that check.

FAQ

Frequently asked questions about Annex III

No, not as of 7 October 2026. On 19 May 2026 the Commission published a draft under Article 6(5) of the AI Act. The targeted consultation ran until 23 July 2026, and the Commission plans to adopt the final version by the end of 2026. Until then, the draft is a useful interpretative aid but not binding.
Annex III of the AI Act (Regulation (EU) 2024/1689) lists specific uses in eight areas for which AI systems count as high-risk under Article 6(2): (1) biometrics, (2) critical infrastructure, (3) education and vocational training, (4) employment, workers management and access to self-employment (including screening and filtering applications), (5) essential private and public services, (6) law enforcement, (7) migration, asylum and border control, and (8) the administration of justice and democratic processes. The specific intended use decides. By way of exception, such a system is not high-risk if Article 6(3) applies; the provider must document this assessment before placing the system on the market or putting it into service (Article 6(4)). The exemption never applies where the system profiles natural persons. SimpleAct’s rule-based risk classification points to Annex III and records the reasoning; a person decides.
Annex III contains eight areas. Below them are, by our count, 25 individual use cases – from remote biometric identification (point 1(a)) to influencing elections (point 8(b)). The Commission can amend and extend the list by delegated act under Article 7.
No. Under Annex III point 4, systems are high-risk if they analyse, filter or evaluate applications, place targeted job ads, or co-determine promotion, termination, task allocation and performance evaluation. AI that only drafts job descriptions from input or coordinates appointments falls under the Article 6(3) exemption according to the draft guidelines – but must be documented and registered.
Since the Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026), the obligations for Annex III high-risk AI apply from 2 December 2027 instead of 2 August 2026 as originally planned. For AI in products under Annex I, the date is 2 August 2028.
An Annex III system is not high-risk if it poses no significant risk to health, safety or fundamental rights because it performs a narrow procedural task, improves a completed human result, detects patterns or deviations without replacing human assessment without review, or performs a preparatory task. If the system profiles natural persons, it is always high-risk.
Yes. The provider must document the assessment before placing the system on the market (Art. 6(4)) and register it in the EU database under Article 49(2). The Digital Omnibus kept this and provided a simplified procedure. The documentation must be provided to the authority on request.
The provider is responsible for the classification and the Article 6(3) assessment. Deployers should still understand it, because their Article 26 duties depend on it. Anyone who uses a system for a purpose it was not intended for, making it high-risk, may become a provider under Article 25.
Annex I is linked to products: AI is high-risk if it is a safety component of a product under EU harmonisation legislation (such as medical devices or machinery) and third-party assessment is required. Annex III is linked to the intended purpose. The Article 6(3) exemption only exists for Annex III. Obligations apply from 2 December 2027 for Annex III and from 2 August 2028 for Annex I.
Not as such. General-purpose AI models have their own rules. For Annex III, the concrete use decides: if a company uses a language model to score applications or rate credit requests, a high-risk system can result – and depending on the set-up, the company can become the provider.
Annex III point 5(b) covers the creditworthiness of natural persons. Rating the creditworthiness of companies does not fall under this point. Take care with sole traders and the self-employed, where the assessment often relates to the person. Systems for detecting financial fraud are expressly excluded.
High-risk AI placed on the market or put into service before the relevant date generally only has to meet the requirements if it is significantly changed afterwards. Systems intended for use by public authorities must comply by 2 August 2030 at the latest.
Breaches of the high-risk obligations can be fined under Article 99(4) up to €15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs, including start-ups, the lower of the two amounts applies (Art. 99(6)). Since the Digital Omnibus, this also applies to small mid-caps (Art. 3(14b), Art. 99(6a)).
Often yes. Annex III systems frequently assess people systematically and process personal data at scale – typical triggers for a data protection impact assessment under Article 35 GDPR. Deployers should use the information in the instructions for use for this (Art. 26(9) AI Act).

More questions? We're happy to help. Send email · Try 7 days (credit card required)

Sources and status

Last updated: · SimpleAct editorial team

This page paraphrases Annex III and gives first orientation. It is not legal advice and does not replace a case-by-case review. As of 7 October 2026, the Commission guidelines were only available as a draft dated 19 May 2026.

Changes

  • 30 October 2026 – New answers to frequently asked questions added.
  • 2026-10-07 – Guidelines section: status checked against the Commission page (still a draft only), key points with paragraph numbers and FAQ added.
  • 2026-10-03 – Fully revised: all 25 use cases with example and counter-example, Article 6(3) check and template, obligations table, draft guidelines.
  • 2026-07-30 – Deadlines updated after the Digital Omnibus.
  • 2026-03-01 – First published.

Document your Annex III classification instead of guessing

In SimpleAct you record every AI system, map it to Annex III, document the Article 6(3) assessment and keep obligations and deadlines in view – traceable for audits and authorities.