Which Systems Fall Into Annex III High Risk Categories

The EU AI Act defines high risk AI systems through Annex III categories, and understanding whether your system falls into this list is the foundation for compliance planning. Annex III sets out ten categories of AI systems that are presumed to pose a high risk to fundamental rights and safety.

The first two categories focus on safety-critical applications. Biometric systems that are not based on prohibited practices fall into the high risk classification. This includes facial recognition systems used for identification purposes in law enforcement, critical infrastructure monitoring through video analytics, and other biometric identification systems. The second category covers AI systems in employment and worker management contexts. These include systems that assess, monitor or evaluate employees or job candidates for recruitment, promotion, performance monitoring or termination decisions. A résumé screening tool that flags candidates or a system that monitors worker productivity and recommends disciplinary action both belong here.

Education and training systems form another high risk cluster. Any AI system that determines access to or the assignment of individuals to educational institutions or training programmes is classified as high risk. This includes admissions systems at universities, course recommendation engines and systems that evaluate learner progress to assign students to different educational pathways.

Critical infrastructure management systems represent a fourth category. AI systems used to operate or manage critical infrastructure in energy, water, transport or healthcare fall into high risk. A system that controls electricity distribution decisions, manages water treatment plant operations or routes emergency vehicles would qualify.

Law enforcement and criminal justice systems carry high risk designation. AI systems used by law enforcement to detect, investigate or prosecute criminal offences include facial recognition for identifying suspects, predictive policing systems, and tools that analyse evidence. Judicial systems that support decisions in legal proceedings are also captured.

Migration, asylum and border control systems are designated high risk. This includes systems that assess asylum applications, detect irregular border crossings, or make decisions affecting visa or residence status determination.

The final Annex III categories cover essential public services and civic participation. AI systems that determine eligibility for public benefits and government services, such as housing allocations, welfare payments or social assistance, are high risk. Also, AI systems used to support civic participation, including voting systems and systems that support public consultation or political campaigns, fall within the scope.

The key question is whether your organisation operates any system matching these categories. A financial services company deploying an AI model to assess creditworthiness must determine if that system fits the category covering AI systems used to determine credit or insurance eligibility. A healthcare organisation using an AI tool to triage patients against emergency department resources must consider whether this is a safety-critical system under the critical infrastructure or healthcare provisions.

Article 6(3) introduces a crucial qualifier: an AI system in Annex III can be excluded from high risk classification if, after a risk assessment and impact assessment, the organisation determines that the system would not pose a significant risk to fundamental rights. This filter is not a blanket exemption. An employer must still demonstrate through impact assessment why their biometric system for workplace access control poses no significant risk. A university must show why their course recommendation engine does not create substantial rights or safety risks.

The temporal dimension matters. Regulation (EU) 2026/1744, adopted in July 2026, deferred full compliance for Annex III high risk systems to 2 December 2027. Before that date, some obligations already apply through Articles 6, 8 and 26, but the full documentation, quality management and conformity assessment regime takes effect from 2 December 2027. For organisations with systems already in operation or under development, this means compliance planning must begin now.

Classification is not a one-time exercise. As your system evolves, the technical design changes, the data characteristics shift or the usage context expands, you must revisit whether your system remains high risk. A tool that starts as an internal employee training recommendation system might later be exposed to customers, changing its risk profile.

The compliance burden for high risk systems is substantial. From 2 December 2027, organisations must have completed technical documentation, risk assessments, impact assessments, quality management systems, conformity assessment and CE marking. This timeline is firm. Organisations should begin classification and compliance planning immediately. Waiting until late 2027 to start assessment and documentation will result in non-compliance.

Understanding the Annex III categories is the first step, but it is not the end. Each category has nuances. A biometric system used only internally for employee badges might present different risks than one used for public identification. An employment system used in hiring presents different considerations than one used for performance management. Understanding these distinctions helps assess whether your specific system, in your specific context, truly falls within the high risk classification. The assessment is not abstract but grounded in your organisational reality.

Frequently asked questions

What is a high risk AI system under the EU AI Act?

A high risk AI system is an AI system listed in Annex III of the EU AI Act that poses risks to fundamental rights or safety. These include biometric identification systems used by law enforcement, AI systems affecting employment and worker management, systems in education and training, critical infrastructure systems, law enforcement tools, migration and asylum systems, and systems determining eligibility for public services. Systems in these categories are subject to complete compliance requirements under Articles 8 to 27.

When do EU AI Act high risk obligations apply?

Most high risk obligations apply from 2 December 2027 under the deferral introduced by Regulation (EU) 2026/1744. However, certain Articles (6, 8 and 26) apply immediately and govern classification, risk assessment and deployer duties. Article 10 data governance and Article 14 human oversight should be implemented in advance to be ready for the December 2027 complete compliance deadline.

What is Annex III of the EU AI Act?

Annex III is the list of AI system categories presumed to pose high risk to fundamental rights and safety. It includes ten categories covering biometric systems, employment systems, education systems, critical infrastructure, law enforcement, criminal justice, migration and asylum, and systems determining eligibility for public services or civic participation.

What is a fundamental rights impact assessment?

A fundamental rights impact assessment (FRIA) is a systematic assessment of how an AI system might affect fundamental rights and freedoms including privacy, non-discrimination, freedom of expression and fair treatment. Public bodies deploying high risk systems must conduct FRIAs before deployment and review them when systems change significantly. The FRIA identifies potential risks, evaluates their probability and severity and documents mitigations.

Do high risk AI systems need CE marking?

Yes, high risk AI systems must be CE marked after successful conformity assessment. CE marking indicates that the system has been assessed for conformity with EU AI Act requirements. The mark is applied to the product or its documentation. CE marking is not a quality guarantee but evidence that conformity assessment was completed and passed.

What is conformity assessment under the EU AI Act?

Conformity assessment is the process whereby a Competent Conformity Assessment Body reviews technical documentation, risk assessments, quality management systems and compliance evidence to determine whether a high risk system meets EU AI Act requirements. The body may conduct a Stage 1 planning meeting and Stage 2 detailed evaluation, issuing a conformity certificate if the system passes.

What technical documentation does a high risk AI system need?

Technical documentation must include a general description of the system, its intended purpose and decision logic, performance characteristics broken down by demographic groups, training and testing data descriptions including representativeness analysis, instructions for use, risk management procedures, quality management system details, human oversight procedures, monitoring and logging arrangements and model transparency information. This documentation is specified in Article 11 and Annex IV.

Can an Annex III system be exempt from high risk classification?

Yes, Article 6(3) allows exclusion from high risk classification if the organisation conducts risk and impact assessment demonstrating that the system would not pose a significant risk to fundamental rights and safety. However, this is not an automatic exemption. The organisation must provide evidence through assessment that residual risk is insignificant and that this assessment is documented.

What are deployer obligations for high risk AI systems?

Deployer obligations under Article 26 require that organisations using high risk systems use them according to instructions, assign qualified human oversight with authority and training, monitor performance through logging, control input data quality, inform affected individuals that AI is involved, maintain records for appropriate periods and report serious incidents to providers and authorities. Public bodies deploying high risk systems must also conduct fundamental rights impact assessments under Article 27.