EU AI Act High-Risk AI Systems: Are You Affected?

The EU AI Act’s high-risk category is where most compliance work happens. If your organization uses AI in employment, credit, education, or several other domains affecting EU residents, you are likely operating high-risk AI systems that require a full compliance program.

What makes an AI system high-risk

The EU AI Act classifies AI systems as high-risk based on their domain and intended purpose, not their technical sophistication. The key question is: does this AI system, when deployed, pose a significant risk to health, safety, or fundamental rights?

The Act uses two lists. Annex I covers AI systems that are safety components of products already regulated under existing EU law (medical devices, machinery, aviation). Annex III covers AI systems in eight specific domains where the potential for harm to individuals is high.

For most businesses, Annex III is the relevant list.

The eight high-risk categories under Annex III

Category 1: Biometric identification and categorization

AI systems used for the remote biometric identification of individuals fall here, with the exception of real-time identification in public spaces (which is mostly prohibited). This includes AI used to verify identity through facial recognition in access control or authentication systems.

Biometric categorization systems that classify individuals based on biometric data into categories like gender, ethnicity, or political opinion are high-risk.

Category 2: Critical infrastructure management

AI systems intended to be used as safety components in the management and operation of road, rail, water, or air transport infrastructure. Electricity, gas, heating, or water supply. And digital infrastructure.

Category 3: Education and vocational training

AI systems used to determine access to, or to direct individuals through, educational institutions. AI systems that evaluate learning outcomes, assess skills for qualification purposes, or monitor and detect prohibited behavior during exams.

This category affects universities, professional training organizations, and companies using AI to screen candidates’ qualifications.

Category 4: Employment, worker management, and access to self-employment

This category directly affects most businesses that use AI in HR. Covered systems include:

• AI used to recruit or select individuals, especially to advertise job vacancies, screen or filter applications, and evaluate candidates during interviews or assessments
• AI used to make decisions on promotion and termination of employment relationships
• AI used to allocate tasks, monitor and evaluate performance, determine pay, and manage work

If your organization uses AI-assisted resume screening, interview AI, performance management AI, or workforce scheduling AI that affects individual workers, it falls under this category.

Category 5: Access to essential private services and public services and benefits

AI systems used by natural or legal persons to evaluate creditworthiness or establish credit scores, AI systems used to assess risk and price in health and life insurance, and AI systems used to manage applications for public social assistance benefits.

Financial services and insurance companies using AI in underwriting, credit scoring, or claims processing are significantly affected by this category.

Category 6: Law enforcement

AI used by law enforcement for risk assessment, polygraph evaluation, evidence reliability assessment, predicting criminal behavior, and profiling. This primarily affects public sector organizations but also private security in some contexts.

Category 7: Migration, asylum, and border control

AI systems used in migration management, including lie detectors, risk assessment for irregular migration, and automated processing of applications for visa, asylum, or residence permits. Primarily affects government entities.

Category 8: Administration of justice and democratic processes

AI used to assist judicial authorities in researching, interpreting, and applying the law. AI used to influence elections or democratic processes. Primarily affects the public sector and political organizations.

Are your AI systems affected?

Most businesses need to focus on Categories 3, 4, and 5, which cover the AI applications most common in corporate use: hiring AI, HR management AI, and financial services AI.

Questions to determine if your AI is in scope:

• Does your AI system influence, assist in, or make decisions about hiring, promotion, termination, or pay?
• Does your AI system evaluate candidates, employees, or contractors?
• Does your AI system assess creditworthiness, insurance risk, or loan eligibility?
• Does your AI system determine access to educational programs or evaluate educational performance?
• Does your AI system process biometric data to identify or categorize individuals?

If yes to any of these for EU residents, you have high-risk AI systems under the Act.

Compliance requirements for high-risk systems

High-risk AI systems must meet seven requirements, all of which must be in place before deployment.

Risk management system. A documented, iterative process for identifying, evaluating, and mitigating risks throughout the system’s lifecycle.

Data and data governance. Documentation of training and testing data, including provenance, quality criteria, and compliance with data protection requirements.

Technical documentation. Comprehensive documentation enabling authorities to assess compliance with all requirements.

Record-keeping. Automatic logging of the system’s operation at a level sufficient for post-market surveillance.

Transparency and provision of information to users. Clear information for deployers and users about the system’s purpose, capabilities, limitations, and the human oversight measures in place.

Human oversight. Design measures that allow effective human oversight, including the ability to understand, monitor, and intervene in the system’s operations.

Accuracy, robustness, and cybersecurity. Performance meets defined accuracy thresholds and the system is secure against AI-specific threats.

The conformity assessment process

Before deploying a high-risk AI system in the EU, providers must complete a conformity assessment.

For most Annex III systems, providers may conduct a self-assessment using the internal control procedure. The provider evaluates their system against the requirements, documents the assessment, and signs an EU Declaration of Conformity.

For biometric identification systems, a third-party conformity assessment by a notified body is required. Notified bodies are accredited organizations designated by EU member states.

After a successful conformity assessment, the system must be registered in the EU AI database before deployment.

Third-party AI services and liability

Most businesses are deployers rather than providers: they use AI built by someone else. Deployers still have obligations under the Act.

Deployers must use high-risk AI systems in accordance with the provider’s instructions, ensure human oversight, monitor performance, and report incidents. They cannot simply point to the vendor’s compliance as a substitute for their own.

When deployers adapt or extend a high-risk AI system in ways beyond its intended purpose, they may take on provider obligations for those modifications. Review vendor contracts carefully to understand what obligations remain with the provider.

For the full EU AI Act overview, see EU AI Act explained. For a compliance checklist, see EU AI Act compliance checklist.

Frequently asked questions

If I use an HR software platform that has AI features, am I a deployer of high-risk AI?

Likely yes. If the HR platform uses AI to assist in screening candidates, evaluating employee performance, or managing employment relationships, those AI features fall under Category 4 (employment and worker management). You are a deployer, and you have deployer obligations under the Act, including ensuring human oversight and monitoring for compliance.

What is the threshold for “significant impact” that determines high-risk classification?

The Act does not use a strict quantitative threshold. The classification is based on the domain and intended purpose. If your AI system falls within the eight Annex III categories by purpose and domain, it is high-risk regardless of how often its outputs are actually significant. The category is defined by potential impact, not average impact.

Can we deploy a high-risk AI system while the compliance program is still being built?

Technically, no. The Act requires compliance before deployment. In practice, many organizations have AI systems already deployed that would be classified as high-risk. For existing systems, build the compliance program as rapidly as possible, prioritize the most significant gaps, and document your remediation timeline. Engage legal counsel on your specific regulatory exposure.

Do you have high-risk AI systems that need a compliance program?

Most organizations with AI in HR, finance, or education have high-risk AI systems. The question is whether those systems have the documentation, oversight, and monitoring the Act requires.

Path one: run a compliance assessment. An AI audit maps your AI systems against EU AI Act risk classifications and identifies which systems need a compliance program and what that program requires.

Path two: work with Phos AI Labs. If you want expert help building high-risk AI system compliance programs for your organization, Phos AI Labs is a CCA-F certified Claude implementation partner. Thirty minutes, no deck. Start here.