EU AI Act High-Risk Obligations

Obligations applying to high-risk AI systems under the EU AI Act. High-risk classification triggers extensive compliance work: risk management, data governance, technical documentation, record-keeping, transparency, human oversight, accuracy/robustness/cybersecurity, conformity assessment, CE marking, post-market monitoring, serious incident reporting. Applicable from 2 August 2026 (Annex III) and 2 August 2027 (Annex I products).

Who must comply

Three primary roles:

• Provider — natural or legal person developing or having developed an AI system, placing it on the market or putting it into service under own name or trademark.
• Deployer — natural or legal person using an AI system under its authority (other than personal non-professional activity). Was called "user" in earlier drafts.
• Importer / Distributor — for AI systems imported into or distributed within the EU.

Additional roles: Authorized representative (for non-EU providers), Product manufacturer (for Annex I product cases).

Each role has distinct obligations; provider obligations are the heaviest.

Provider obligations (Articles 16, 8-21)

Risk management system (Article 9)

Establish and maintain a risk management system throughout the AI system lifecycle. Iterative process including:

• Identification and analysis of risks the system may pose to health, safety, fundamental rights
• Estimation and evaluation of risks that may emerge when used as intended
• Evaluation of risks based on post-market monitoring data
• Adoption of risk management measures

Data and data governance (Article 10)

For high-risk systems using training/validation/testing data:

• Data must be subject to appropriate data governance and management practices
• Data sets must be relevant, sufficiently representative, free of errors (to extent feasible), complete in view of intended purpose
• Special attention to bias in training data
• Special-category personal data processed only when strictly necessary for bias monitoring/correction, with appropriate safeguards

Technical documentation (Article 11, Annex IV)

Comprehensive technical documentation including:

• General description of the system, intended purpose, lifecycle
• Detailed system description (elements, processes, computational logic, optimization)
• Detailed monitoring, functioning, control description
• Risk management system
• Changes through lifecycle
• Lists of harmonized standards applied or solutions adopted to meet requirements
• EU declaration of conformity
• Description of system in use

Annex IV provides the documentation template (extensive).

Record-keeping (Article 12)

High-risk AI systems must enable automatic logging during operation. Logging:

• Must be appropriate to intended purpose
• Must support functioning monitoring
• Must support post-market monitoring

Logs must be kept by the provider when provider deploys the system, or by the deployer.

Transparency and information for users (Article 13)

Systems designed and developed so deployers can:

• Understand and use the system appropriately
• Receive concise / complete / correct / clear information

Instructions for use must include: identity and contact of provider, characteristics, capabilities, limitations, intended purpose, accuracy / robustness / cybersecurity level, foreseeable circumstances of malfunction, expected useful life and maintenance.

Human oversight (Article 14)

Systems designed and developed so they can be effectively overseen by natural persons. Human oversight measures include:

• Ability to fully understand the capacities and limitations
• Awareness of automation bias risks
• Ability to correctly interpret outputs
• Ability to disregard, override, or reverse outputs
• Ability to intervene or stop the system

Specific Article 14(5) provisions for certain biometric-identification high-risk systems requiring "two-eye" verification.

Accuracy, robustness, cybersecurity (Article 15)

System must be designed and developed to achieve:

• Appropriate level of accuracy
• Robustness against errors, faults, inconsistencies
• Resilience against unauthorized third parties (cybersecurity)
• Mitigation against attacks attempting to manipulate training data or alter model behavior

Accuracy levels and metrics declared in instructions for use.

Quality management system (Article 17)

Provider shall put in place a quality management system to ensure compliance. Documented in policies, procedures, instructions. Includes:

• Strategy for regulatory compliance
• Techniques for design / quality control / verification
• Procedures for conducting tests
• Procedures for post-market monitoring
• Procedures for serious incident reporting
• Procedures for cooperation with authorities

Conformity assessment (Articles 19, 43)

Before placing on market or putting into service:

• Internal control conformity assessment for most Annex III high-risk systems (provider self-assesses).
• Notified Body assessment for some categories (notably biometric remote identification subject to specific procedures, and Annex I product cases following the underlying product regulation's conformity assessment).

Conformity assessment results in EU declaration of conformity (Annex V) and CE marking.

Registration in EU database (Article 49)

High-risk AI systems must be registered in an EU database before being placed on market or put into service.

Post-market monitoring (Article 72)

Provider must establish and document post-market monitoring system to actively collect, document, analyze data on performance throughout the lifetime. Outputs feed back into risk management.

Reporting of serious incidents (Article 73)

Providers must report serious incidents to market surveillance authority. Timelines:

• Within 15 days of becoming aware of serious incident or malfunction
• Shorter timeline (immediately, no later than 2 days) for incidents involving breach of fundamental rights obligations
• Even shorter (immediately, no later than 10 days) for incidents involving widespread infringement / serious malfunction

CE marking (Article 48)

CE marking affixed to the high-risk AI system (or accompanying documentation where physical attachment not feasible).

Cooperation with authorities (Article 21)

Providers cooperate with national competent authorities and market surveillance authorities.

Deployer obligations (Article 26)

For deployers of high-risk AI systems:

• Use system according to instructions for use
• Assign human oversight to natural persons with necessary competence
• Monitor operation per instructions
• Suspend use if reasonable grounds to consider use makes the system present risk
• Inform provider, distributor, market surveillance authority of risks identified
• Keep logs automatically generated by the system (for at least 6 months unless other rules apply)
• For workplace deployment: inform affected workers and their representatives prior to use
• Conduct fundamental rights impact assessment in certain cases (Article 27, deployer use of high-risk by public authorities and private entities providing public services)

Specific deployer obligations: Fundamental rights impact assessment (Article 27)

For deployers of high-risk Annex III systems (some public/quasi-public uses), additional obligation: conduct FRIA before deploying:

• Process description
• Period of use
• Categories of natural persons / groups likely affected
• Specific risks of harm
• Implementation of human oversight measures
• Measures if risks materialize, including governance / complaint mechanisms

Result registered with market surveillance authority in some cases.

Importer / distributor obligations (Articles 23-25)

• Importer (Article 23): verify provider compliance, ensure documentation accompanies the system, indicate own contact information, monitor for compliance signals, cooperate with authorities.
• Distributor (Article 24): verify CE marking and required documentation, ensure storage / transport conditions don't compromise compliance, cooperate with authorities.

Authorized representative (Article 22)

Non-EU providers must designate an EU-established authorized representative who:

• Verifies that conformity assessment and technical documentation are completed
• Keeps documentation available for authorities
• Cooperates with authorities
• Acts as point of contact

Notified Body involvement

For high-risk AI systems requiring third-party conformity assessment:

• Notified Body designated by Member State for the specific category
• Performs conformity assessment per Annex VI or Annex VII procedure
• Issues conformity certificate
• Subsequent surveillance during validity

Notified Body capacity is a current bottleneck (as of 2026); designations are ongoing through 2025-2027.

Compliance evidence per obligation

SRE and AI-agent fit notes

Agent systems and high-risk classification

Most agent systems built on model APIs are not in Annex III by default. Use case matters:

• Agent helping HR with hiring → Area 4 (Employment) → likely high-risk
• Agent making credit decisions → Area 5 (Essential services) → likely high-risk
• Agent moderating content for major platforms → potentially Area 8 edge
• Internal-productivity agent without consequential decisions → likely not high-risk

Article 6(3) exemption applies if the system performs narrow procedural tasks or doesn't pose significant risk. Document the assessment.

Mapping high-risk obligations to existing implementation work

For systems crossing into high-risk territory, obligations map heavily to existing AI governance work:

ISO 42001 harmonization (when complete) will provide presumption of conformity for many of these obligations.

Stefan-context implementation sketch

For solo / small-team agent system work potentially touching high-risk categories:

• Run classification analysis per significant agent.
• Most likely outcome: limited-risk or minimal-risk; document the Article 6(3) exemption rationale where applicable.
• For systems crossing into high-risk: coordinate with client; substantial implementation work; consider ISO 42001 alignment path.
• Document classification assessments in vault as evidence artefacts.

See also

• cluster MOC · EU AI Act Risk Tiers · EU AI Act Timeline · EU AI Act GPAI Obligations · EU AI Act Governance
• ISO 42001 Clause Structure · ISO 42001 Annex A Controls · NIST AI RMF Core Functions · OWASP LLM Top 10 2025