High-Risk AI Systems Under the EU AI Act: Complete Guide

Updated 25 May 2026 after the EU's Digital Omnibus agreement (7 May 2026): the Annex III high-risk compliance deadline has moved from 2 August 2026 to 2 December 2027 (Annex I to 2 August 2028). Article 50(2) transparency obligations still apply 2 August 2026. See our Omnibus update for the full timeline.

Why High-Risk Classification Matters

The EU AI Act's compliance obligations are concentrated almost entirely on high-risk AI systems. Minimal-risk systems face no mandatory requirements. Limited-risk systems need only transparency disclosures. But high-risk systems must meet a comprehensive set of requirements covering documentation, risk management, data governance, human oversight, accuracy, cybersecurity, conformity assessment, registration, and ongoing monitoring.

Getting the classification right is therefore the most consequential step in EU AI Act compliance. Misclassifying a high-risk system as limited risk leaves you exposed to penalties of up to €15 million or 3% of global annual turnover. Misclassifying a limited-risk system as high risk wastes resources on unnecessary compliance work. To skip ahead, you can classify your system using the same Article 5, Annex III, and Article 6(3) decision tree this guide unpacks — three minutes, no signup.

Two Pathways to High-Risk

The AI Act establishes two independent pathways for high-risk classification. A system only needs to trigger one.

Pathway 1: Safety Component (Article 6(1))

An AI system is high-risk if both conditions are met:

1. The AI system is used as a safety component of a product, or is itself a product, covered by EU harmonisation legislation listed in Annex I of the AI Act
2. That product is required to undergo a third-party conformity assessment under the applicable harmonisation legislation

Annex I covers existing EU product safety legislation including:

• Machinery Regulation (EU) 2023/1230
• Medical Devices Regulation (EU) 2017/745
• In-Vitro Diagnostic Medical Devices Regulation (EU) 2017/746
• Radio Equipment Directive 2014/53/EU
• Civil Aviation Regulation (EU) 2018/1139
• Motor Vehicle Type-Approval Regulations
• Marine Equipment Directive 2014/90/EU
• Rail Interoperability Directive (EU) 2016/797
• And others

Examples: An AI system that controls braking decisions in an autonomous vehicle. An AI-powered diagnostic tool classified as a medical device. An AI component managing safety-critical functions in industrial machinery.

Application date: August 2, 2027 (one year later than Pathway 2).

Pathway 2: Annex III Listed (Article 6(2))

An AI system is high-risk if it falls into one of the use case categories listed in Annex III of the AI Act. This is the pathway that applies to most business AI systems.

Application date: August 2, 2026.

The Eight Annex III Categories

Category 1: Biometrics

Sub-area	Description	Examples
1(a)	Remote biometric identification	Post-hoc facial recognition, identification in non-public spaces
1(b)	Biometric categorization by sensitive attributes	Categorizing people by ethnicity, gender, or other protected characteristics from biometric data
1(c)	Emotion recognition	Inferring emotions from facial expressions, voice patterns, body language (outside the workplace/education ban under Article 5)

Note: Real-time biometric identification in public spaces for law enforcement is not high-risk — it is prohibited under Article 5(1)(h), with three narrow exceptions requiring judicial authorization. Emotion recognition in the workplace and education is also prohibited (Article 5(1)(f)), except for medical or safety purposes.

Category 2: Critical Infrastructure

Sub-area	Description	Examples
2(a)	Safety components in management/operation of critical digital infrastructure, road traffic, water/gas/heating/electricity supply	AI managing electricity grid load balancing, AI controlling traffic light systems, AI monitoring water treatment processes

The system must be a safety component — not just any software used by a utility or infrastructure provider. An AI-powered billing system at an energy company is not high-risk. An AI system making decisions about power distribution is.

Category 3: Education and Vocational Training

Sub-area	Description	Examples
3(a)	Determining access, admission, or assignment to educational institutions	AI-powered university admissions, automated school placement
3(b)	Evaluating learning outcomes, including steering the learning process	AI grading systems, adaptive learning platforms that determine curriculum
3(c)	Assessing appropriate education level and influencing access to education	AI that decides what level of education someone should receive
3(d)	Monitoring and detecting prohibited behavior during tests	AI exam proctoring, cheating detection systems

Category 4: Employment, Worker Management, Access to Self-Employment

Sub-area	Description	Examples
4(a)	Recruitment and selection: targeted job ads, screening/filtering applications, evaluating candidates	AI-powered CV screening, automated interview scoring, AI job ad targeting
4(b)	Decisions on work relationships: promotion, termination, task allocation based on behavior/traits; monitoring and evaluating performance	AI-driven performance reviews, automated task assignment based on employee profiling, AI-powered workforce management

This category is particularly relevant for SMEs. Any AI tool used in hiring or employee evaluation falls here — including third-party SaaS tools. If you use an AI-powered applicant tracking system, you are a deployer of a high-risk AI system.

Category 5: Access to Essential Services

Sub-area	Description	Examples
5(a)	Evaluating eligibility for public assistance, benefits, or services	AI determining welfare eligibility, healthcare access decisions
5(b)	Evaluating creditworthiness or establishing credit scores	AI credit scoring, automated lending decisions. Exception: fraud detection is explicitly not high-risk
5(c)	Risk assessment and pricing for life and health insurance	AI determining insurance premiums for natural persons
5(d)	Evaluating and classifying emergency calls, triaging emergency patients	AI-powered 112/911 dispatch, emergency room triage systems

The fraud detection exception in 5(b) is notable. Banks using AI solely for fraud detection do not trigger high-risk classification for that system, even though credit scoring and lending AI does.

Category 6: Law Enforcement

Sub-area	Description	Examples
6(a)	Assessing risk of person becoming crime victim	Victim risk assessment tools
6(b)	Polygraphs and lie detectors	AI-based deception detection
6(c)	Evaluating reliability of evidence	AI analyzing evidence in criminal investigations
6(d)	Assessing risk of offending or re-offending	Recidivism prediction (must augment human assessment, not replace it)
6(e)	Profiling during detection, investigation, or prosecution	Criminal profiling AI

Category 7: Migration, Asylum, and Border Control

Sub-area	Description	Examples
7(a)	Polygraphs and lie detectors in migration context	AI deception detection at borders
7(b)	Assessing irregular migration risk or health risk	Risk assessment for persons intending to enter the EU
7(c)	Examining asylum, visa, or residence applications	AI processing immigration applications
7(d)	Detecting, recognizing, or identifying persons in migration context	Exception: travel document verification

Category 8: Administration of Justice and Democratic Processes

Sub-area	Description	Examples
8(a)	Assisting judicial authorities in interpreting and applying law	AI legal research tools used by courts, AI in alternative dispute resolution
8(b)	Influencing election outcomes or voting behavior	Exception: AI for organizing political campaigns that does not directly interact with voters

The Article 6(3) Exception: When High-Risk Can Be Downgraded

Not every Annex III system is automatically high-risk. Article 6(3) provides an exception pathway that can downgrade a system that technically falls within an Annex III category.

The exception applies when a system meets both conditions:

Condition 1: Limited Function

The system performs only one of the following:

• (a) A narrow procedural task (e.g., data format conversion, document sorting)
• (b) Improving the result of a previously completed human activity (e.g., spell-checking a human-written assessment)
• (c) Detecting patterns for human review without replacing or influencing the human assessment (e.g., anomaly flagging)
• (d) A preparatory task for a human assessment (e.g., data gathering before a human makes the decision)

Condition 2: No Profiling

The system does not profile natural persons. The GDPR Article 4(4) definition of profiling applies: automated processing of personal data to evaluate personal aspects such as work performance, economic situation, health, preferences, reliability, behavior, location, or movements.

If profiling is involved, the system is always high-risk — the Article 6(3) exception cannot apply, regardless of the function.

Exception Obligations

Even when the exception applies and the system is downgraded, the provider must:

1. Document the assessment explaining why the exception applies
2. Register the system in the EU database with the documented reasoning (Article 49(2))
3. Notify the relevant market surveillance authority

This is not a "get out of compliance free" card. It is a documented exception with its own obligations.

Provider Obligations for High-Risk Systems

Providers — those who develop high-risk AI systems or place them on the market under their name — bear 19 distinct obligations:

#	Article	Obligation
1	Art. 9	Establish and maintain a continuous risk management system
2	Art. 10	Implement data governance for training, validation, and testing data
3	Art. 11	Prepare and maintain Annex IV technical documentation
4	Art. 12	Design automatic logging capabilities into the system
5	Art. 13	Provide transparency information and instructions for use
6	Art. 14	Design human oversight measures into the system
7	Art. 15	Ensure accuracy, robustness, and cybersecurity
8	Art. 17	Implement a quality management system
9	Art. 18	Retain documentation for 10 years
10	Art. 19	Retain auto-generated logs for minimum 6 months
11	Art. 20	Take corrective actions for non-compliant systems
12	Art. 21	Cooperate with authorities upon request
13	Art. 22	Appoint an EU authorized representative (if non-EU provider)
14	Art. 43	Complete conformity assessment before market placement
15	Art. 47	Draw up EU declaration of conformity
16	Art. 48	Affix CE marking
17	Art. 49	Register in the EU database
18	Art. 72	Establish post-market monitoring system
19	Art. 73	Report serious incidents within 15 days

For deployers in scope of Article 27, our Article 27 FRIA template covers the required elements out of the box. The end-to-end compliance checklist walks each of these provider obligations in sequence with the article references intact.

Conformity Assessment: Self or Third-Party?

Most high-risk systems undergo self-assessment based on Annex VI. The provider verifies their own compliance, issues a declaration of conformity, and affixes the CE marking.

Third-party assessment by a notified body (Annex VII) may be required for Annex III point 1 biometric systems — remote biometric identification, biometric categorisation, and emotion recognition — where harmonised standards or common specifications do not fully cover the applicable requirements. Other Annex III systems generally use internal control under Annex VI.

Deployer Obligations for High-Risk Systems

Deployers — those who use high-risk AI systems under their authority — have a lighter but still significant set of obligations:

#	Article	Obligation
1	Art. 26(1)	Use the system according to the provider's instructions
2	Art. 26(2)	Assign competent, trained persons for human oversight
3	Art. 26(4)	Ensure input data is relevant and representative
4	Art. 26(5)	Monitor operation, suspend if risks arise, report incidents
5	Art. 26(6)	Retain auto-generated logs for minimum 6 months
6	Art. 27	Conduct a FRIA (only for: public bodies, public service providers, credit scoring deployers, life/health insurance deployers)
7	Art. 26(11)	Inform persons affected by AI-based decisions
8	Art. 26(7)	Inform workers and worker representatives when AI is used in the workplace
9	Art. 50(3)	Inform persons exposed to emotion recognition or biometric categorization

Role Shifting Warning

Article 25 specifies that a deployer becomes a provider — inheriting all 19 provider obligations — if they:

• Put their own name or trademark on a high-risk AI system
• Make a substantial modification to the system
• Change the intended purpose of any AI system in a way that makes it high-risk

This is a critical consideration for companies that customize or rebrand third-party AI tools. The full provider obligation set is not free to produce, which is why our breakdown of EU AI Act compliance costs is worth reading before any role-shift decision.

Timeline: When Do High-Risk Obligations Apply?

Pathway	Application Date	Scope
Annex III (Pathway 2)	August 2, 2026	All 8 categories listed above
Annex I product-embedded (Pathway 1)	August 2, 2027	Safety components of regulated products

There is a transitional provision: high-risk AI systems already on the market before August 2, 2026 need to comply only if they undergo a significant design change after that date. Systems intended to be used by public authorities are not open-endedly grandfathered: providers and deployers must take the necessary steps to comply by August 2, 2030.

Getting Started

Classification is the foundation. If you are unsure whether your AI system is high-risk, the Witness AI System Classifier walks you through the decision tree in about three minutes — covering the Article 5 prohibitions check, Annex III category matching, and the Article 6(3) exception assessment. It identifies your risk level, your role, and the specific obligations that apply.