The EU AI Act entered into force on August 1, 2024. Article 5 prohibited AI provisions applied from February 2025. Title III high-risk AI obligations apply from August 2, 2026. Organizations placing AI systems in the EU market or affecting people in the EU should begin audit preparation now.
This checklist addresses evidence that high-risk and limited-risk AI systems must provide before an audit or conformity assessment. It excludes general-purpose AI model obligations under Title VIIIa, which follow a separate documentation regime. The scope is systems already in your enterprise that fall under Annexes II and III.
Risk Tiers
The EU AI Act uses four tiers. Article 5 prohibits social scoring by public authorities, biometric categorization using sensitive characteristics for discriminatory purposes, real-time remote biometric identification in publicly accessible spaces by law enforcement, with narrow exceptions, and subliminal manipulation. Systems in this category stop the process before the checklist.
Annex II covers high-risk AI products already governed by EU product safety directives. Annex III covers standalone high-risk applications in sensitive areas: biometric identification, critical infrastructure management, educational and vocational training systems determining access or outcomes, employment and HR systems making consequential decisions, access to essential private and public services, law enforcement, migration and asylum management, and administration of justice.
Limited-risk AI includes chatbots and systems that generate synthetic content. Its obligation is transparency, meaning users must know they are interacting with AI. Minimal-risk AI carries no specific obligations. Most compliance work falls on the high-risk tier.
Article 9: Risk Management Documentation
Article 9 calls for a documented risk management system operating throughout the AI system lifecycle. This is not a single assessment. It requires procedures, recurring reviews, and records of decisions about residual risk.
Preparation should collect a risk management plan covering scope and method, hazard identification and risk analysis for the use case, implemented mitigation measures, authorized acceptance of residual risk, and testing and validation results from before deployment. The system must address health, safety, and fundamental rights. A generic cybersecurity assessment does not replace AI-specific analysis of those dimensions.
Articles 10 and 11: Data and Technical Records
Article 10 addresses data and data governance for systems using training, validation, or testing datasets. Records should describe the datasets, collection and processing methods, possible biases and corrective measures, and the governance practices used. When a provider manages data for a procured system, its documentation must be available to the deployer for audit.
Article 11 requires technical documentation before an AI system reaches the market or enters service. Annex IV lists a system description, design specifications and development method, training and testing data information, validation and testing procedures and results, performance metrics and limits, human oversight measures, and monitoring and logging capabilities.
A frequent gap is provider documentation that the deployer has not reviewed, understood, or adapted. The Act assigns duties to providers, those who develop or place systems on the market, and deployers, those using high-risk AI professionally. A deployer must verify that vendor documentation covers its own use case and deployment context.
Article 12: Logs and Records
Article 12 requires high-risk AI systems to support automatic event logging. Logs must show how the system operates and preserve enough information for later review of outputs. Deployers should confirm that systems generate logs and that records are retained under applicable record-keeping rules, at minimum for the period set by sectoral rules, or three years where none applies.
Audit preparation should establish which systems log events, the storage location, retention period, access rights, and whether the format can reconstruct the basis for decisions affecting particular people. Recommendations are more difficult than deterministic decisions because their input-to-output chain may not be easy to document.
Article 14: Human Oversight
Article 14 requires high-risk AI systems to include human oversight that lets a natural person understand and monitor outputs, interrupt or stop operation when necessary, and override or disregard an output. Documentation should describe built-in safeguards and the deployer's procedures for applying oversight.
Auditors will seek proof that oversight works in practice, not merely that a policy exists. Keep procedures stating when and how staff review outputs, training records showing their understanding of system capabilities and limits, and records of occasions when oversight was used. An unused policy with no supporting training records is insufficient.
Article 9 Conformity Assessment Checklist
For high-risk AI systems in Annex III categories without a notified body, providers must complete an internal conformity assessment before market placement. Deployers must apply the technical and organizational measures needed to use the system as documented. Audit readiness should cover:
- Written risk management plan with regular review schedule and completed review records
- Data governance documentation covering training data sources, quality criteria, and bias examination results
- Technical documentation file per Annex IV either received from provider and supplemented for deployment context, or drafted in-house for internally developed systems
- Logging configuration confirmed and retention schedule documented
- Transparency disclosures in user-facing interfaces for limited-risk systems
- Human oversight procedures written, communicated, and evidenced through training records
- Incident reporting procedures addressing the Article 73 requirements for serious incidents
- Post-market monitoring plan for tracking ongoing performance after deployment
The EU AI Act does not demand flawless high-risk AI. It demands documentation, risk management, monitoring, and human oversight, with records for each. Audit problems usually arise not from imperfect performance, but from governance records that differ from actual operation. Findings collect where documentation and deployment reality diverge.