EU AI Act Compliance Checklist

The EU AI Act (Regulation 2024/1689) organises its regulatory requirements around a risk-based classification system with four tiers: unacceptable risk, high risk, limited risk, and minimal risk. This tiered approach is designed to ensure that the most invasive or dangerous AI applications face the strictest requirements, while low-risk applications remain largely unregulated. Understanding which tier your AI system falls into is the critical first step in any compliance programme.

Unacceptable risk AI systems are banned outright. These are the practices listed in Article 5, which include social scoring by public authorities, real-time remote biometric identification in publicly accessible spaces for law enforcement purposes (with narrow exceptions), exploitation of vulnerabilities of specific groups, and emotion recognition in workplaces and educational institutions. The prohibitions on these practices applied from 2 February 2025 — the first provisions of the AI Act to take effect.

High-risk AI systems are subject to extensive requirements but are permitted provided they meet the compliance obligations set out in Chapter III. These include AI systems used in critical infrastructure, education, employment, essential services, law enforcement, migration, and the administration of justice. They also include AI systems that serve as safety components of products covered by existing EU product safety legislation. Limited-risk systems are subject primarily to transparency obligations — informing users that they are interacting with AI. Minimal-risk systems are subject to no mandatory requirements, though voluntary codes of practice are encouraged.

The classification assessment must be conducted for each AI system the organisation deploys, develops, or provides. A single organisation may have systems falling across multiple tiers. The assessment should be documented and should include the reasoning for the classification, the relevant articles and annexes referenced, and any borderline determinations that were resolved. Regulators will scrutinise classification decisions, particularly for systems that were classified below high-risk where the use case might suggest otherwise.

Chapter III of the AI Act establishes the compliance requirements for high-risk AI systems. These requirements apply to both providers (developers) and deployers (users) of high-risk AI systems, though the obligations differ in scope and nature. For providers, the requirements are comprehensive and include risk management, data governance, technical documentation, record-keeping, transparency, human oversight, and accuracy, robustness, and cybersecurity standards.

The risk management system under Article 9 must be established, implemented, documented, and maintained throughout the entire lifecycle of the high-risk AI system. It must include identification and analysis of known and reasonably foreseeable risks, estimation and evaluation of the risks that may emerge when the system is used in accordance with its intended purpose, adoption of appropriate risk management measures, and testing procedures to ensure that the risk management measures are effective. This is not a one-time assessment — it must be a living process that adapts as the system evolves and as the deployment context changes.

Data governance under Article 10 requires that training, validation, and testing datasets are subject to appropriate data governance practices. These practices must cover the design choices for the datasets, data collection processes, data preparation operations (annotation, labelling, cleaning, enrichment), the formulation of relevant assumptions about the information the data is supposed to measure, an assessment of the availability, quantity, and suitability of the datasets, examination of possible biases, and identification of any data gaps that could affect compliance. For organisations using pre-trained foundation models, demonstrating data governance compliance for the training data can be particularly challenging.

Technical documentation under Article 11 must be drawn up before the system is placed on the market or put into service and must be kept up-to-date. The documentation must contain, at minimum, a general description of the system, a detailed description of the development process, detailed information about monitoring, functioning, and control, a description of the risk management system, and a description of the changes made throughout the lifecycle. Annex IV of the regulation provides the exhaustive list of required documentation elements.

Chapter V of the AI Act introduces requirements for General-Purpose AI (GPAI) models — large foundation models capable of performing a wide range of tasks. These requirements apply to providers of GPAI models and became applicable from 2 August 2025. The GPAI provisions are layered: all GPAI models must meet baseline transparency requirements, and GPAI models with systemic risk must meet additional obligations.

The baseline requirements for all GPAI model providers under Article 53 include: drawing up and keeping up-to-date technical documentation of the model, drawing up and keeping up-to-date information and documentation to be made available to downstream providers who integrate the GPAI model into their AI systems, putting in place a policy to comply with Union copyright law (including the text and data mining opt-out under Article 4 of the Copyright Directive), and publishing a sufficiently detailed summary of the training data according to a template provided by the AI Office.

GPAI models classified as presenting systemic risk under Article 51 face additional obligations. A GPAI model is presumed to have systemic risk if it was trained using a total computing power of more than 10^25 FLOPs, or if the Commission designates it as such based on criteria including the number of registered end users, the degree of multi-modality, and benchmarks and evaluations of capabilities. Providers of systemic risk GPAI models must, in addition to baseline requirements: perform model evaluations including adversarial testing, assess and mitigate systemic risks, track and report serious incidents, and ensure an adequate level of cybersecurity protection for the model.

Organisations that deploy GPAI models developed by third parties (e.g., using APIs from foundation model providers) should ensure that their provider has complied with the GPAI obligations. Contractual provisions should require the GPAI provider to share relevant technical documentation and to notify the deployer of material changes to the model. Due diligence on GPAI provider compliance is becoming an essential part of AI procurement processes.

This section provides a structured checklist of compliance actions organised by the phased timeline. Organisations should use this as a starting framework and adapt it to their specific AI portfolio and risk profile.

Phase 1 (effective 2 February 2025 — immediate action required): Conduct a comprehensive audit of all AI systems in use, including third-party services, to identify any prohibited practices under Article 5. Discontinue any AI systems deploying social scoring, subliminal manipulation, vulnerability exploitation, or prohibited biometric identification. Implement an AI literacy programme for staff who develop, deploy, operate, or interact with AI systems. Document all classification decisions and the reasoning behind them.

Phase 2 (effective 2 August 2025 — now in force): If you are a provider of GPAI models, prepare technical documentation including a training data summary. Implement a copyright compliance policy including mechanisms to honour text and data mining opt-outs. If your GPAI model is classified as systemic risk, conduct model evaluations, adversarial testing, and systemic risk assessments. If you deploy third-party GPAI models, verify your provider's compliance and obtain relevant documentation.

Phase 3 (effective 2 August 2026 — preparation required now): For each high-risk AI system, implement a risk management system covering the full lifecycle. Establish data governance practices for training, validation, and testing datasets. Prepare technical documentation meeting Annex IV requirements. Implement logging and record-keeping capabilities. Design and test human oversight mechanisms. Conduct accuracy, robustness, and cybersecurity assessments. Establish post-market monitoring processes. Complete conformity assessments and prepare EU declarations of conformity. For deployers: implement fundamental rights impact assessments (Article 27), establish user notification processes, and ensure human oversight of automated decisions.

Ongoing: Maintain an AI system register with current risk classifications. Review classifications when systems are updated or use contexts change. Monitor regulatory guidance from the European AI Office, national authorities, and standards bodies. Track emerging harmonised standards under standardisation requests to CEN and CENELEC.