Implementing EU AI Act Compliance: Step-by-Step Checklist for Product Teams

This article contains affiliate links. We may earn a commission at no extra cost to you. Full disclosure.

When the EU AI Act enters full enforcement, product teams face penalties of up to 7% of global annual turnover or €35 million—whichever is higher—for deploying non-compliant high-risk systems. Yet a 2024 survey by the AI Governance Center found that only 12% of organisations have a structured compliance roadmap. The gap between regulatory ambition and engineering reality is widening, and the clock is ticking. This article provides a step-by-step checklist grounded in the Act’s actual text, Annexes, and enforcement timelines, drawing on published compliance frameworks from Microsoft, Google, and the European Commission’s own guidance. Each step includes specific documentation requirements, tool recommendations, and measurable benchmarks so your team can move from panic to process.

1. Classifying Your AI System Under the Four Risk Tiers

The first and most consequential step is determining whether your system falls into the prohibited, high-risk, limited-risk, or minimal-risk category. The Act bans eight specific AI practices outright, including social scoring by governments and real-time biometric surveillance in public spaces for law enforcement (with narrow exceptions). For most product teams, the critical question is whether their system qualifies as “high-risk” under Article 6 and Annex III. Annex III lists 8 areas: biometrics, critical infrastructure, education, employment, access to essential services, law enforcement, migration, and administration of justice. A generative AI chatbot used for customer service in banking, for example, becomes high-risk only if it is used to evaluate creditworthiness—not if it merely answers FAQs.

To classify correctly, use the self-assessment template published by the European Commission’s AI Office in July 2024. It includes 23 yes/no questions covering intended purpose, data sensitivity, and decision-making autonomy. For instance, if your system processes biometric data to infer emotions, it is automatically high-risk regardless of sector. If your system is a standalone software component used in a high-risk product, it inherits that classification. Microsoft’s Azure documentation provides a concrete example: their Face API is classified as high-risk when used for identity verification in access control, but limited-risk when used for photo organisation. Document your classification rationale with version and date—this becomes the first entry in your compliance file.

One common pitfall is assuming that all AI systems using machine learning are high-risk. The Act explicitly exempts systems performing narrow procedural tasks, quality improvement, or preparatory data processing (Article 6(3)). For example, a spam filter for email is minimal-risk. A tool that predicts student dropout based on engagement metrics, however, falls under Annex III(c) for education and is high-risk. The European Commission’s FAQ (June 2024) clarifies that “high-risk” is determined by the system’s function, not its technical complexity. Teams should create a risk classification register, updated quarterly, and store it alongside the system’s model card.

2. Building a Risk Management System That Actually Works

Article 9 requires high-risk providers to establish a continuous, iterative risk management process. This is not a one-time checkbox. The process must identify known and foreseeable risks, evaluate residual risk after mitigation, and implement measures that reduce risk to acceptable levels. The standard is not zero risk—it is that risks are “acceptable” in the context of the system’s intended purpose. For example, a medical diagnosis AI may have a 2% misdiagnosis rate, but if that rate is lower than the human baseline and transparently communicated, it may pass. The European Commission’s draft harmonised standard (prEN 17007, expected Q1 2025) recommends using ISO 31000 as a baseline, with specific adaptations for AI.

Concrete steps: Start by listing all possible harms—physical, psychological, economic, discrimination, privacy—and assign a severity and likelihood score using a 5×5 matrix. For each risk, document the mitigation measure: training data augmentation, model pruning, human-in-the-loop thresholds, or explainability outputs. For instance, to mitigate bias in hiring algorithms, you might implement a 0.8 disparate impact ratio threshold (based on the EEOC’s four-fifths rule) and run post-hoc fairness audits monthly. The risk management file must be updated whenever the system undergoes a “substantial modification” (Article 43(4))—defined as a change that alters the intended purpose or significantly increases risk. The AI Office’s guidance (September 2024) gives the example of adding a new language to a translation system: if the language involves a different demographic, it may trigger a re-assessment.

Tools like Holistic AI’s Risk Manager and Credo AI’s Governance Platform automate much of this tracking. Both integrate with MLOps pipelines to flag modifications. Budget for at least 40 hours of legal and engineering time per system annually just for risk documentation, based on early adopter reports from a 2024 pilot by the Dutch Authority for Digital Infrastructure. Do not rely solely on templates; the Act requires that risk management be “continuous and iterative,” meaning you must demonstrate active monitoring, not just a static PDF.

3. Generating Technical Documentation That Passes Audit

Article 11 and Annex IV specify the technical documentation a provider must draw up before placing a high-risk system on the market. This includes the system’s design, development methodology, training data specifications, performance metrics, and the results of any testing. The documentation must be kept for 10 years after the system is placed on the market (Article 18). A practical approach is to adopt the Model Card framework (originally proposed by Google in 2019) and extend it with the fields required by Annex IV. Hugging Face’s model card template now includes a dedicated “EU AI Act Compliance” section that maps each field to the relevant article.

Specific fields you must include: a detailed description of the system’s intended purpose (including its environment of use), the technical specifications of the training data (size, sources, labeling procedures, cleaning steps), the metrics used to measure accuracy and robustness (e.g., F1 score, AUC-ROC, adversarial robustness against PGD attacks), and the results of bias testing (e.g., demographic parity difference, equal opportunity difference). For example, if you deploy a BERT-based resume screener, you must document the training data’s geographic distribution, the proportion of gender-annotated samples, and the achieved equal opportunity difference of 0.03 (a typical acceptable threshold).

Annex IV also requires a description of the system’s architecture, including the number of layers, parameter count, and training compute (FLOPs). For a large language model, you might report 7 billion parameters trained on 1 trillion tokens using 1,024 A100 GPUs for 14 days—specific numbers matter. The European Commission’s Joint Research Centre published a recommended template in July 2024 that includes a “technical documentation checklist” with 47 items. Teams should version-control this documentation alongside the model weights, using platforms like DVC or MLflow’s model registry. Missing documentation is the most common reason for conformity assessment failures, according to a 2023 simulation by the German Federal Office for Information Security.

4. Implementing Bias Testing and Fairness Monitoring That Holds Up

Article 10 requires that training, validation, and testing data sets are relevant, representative, free from errors, and complete in view of the intended purpose. This is the legal basis for bias testing. The Act does not mandate a specific fairness metric, but the Article 29 Working Party’s guidelines (adopted June 2024) recommend using at least two metrics: one group-based (e.g., demographic parity or equal opportunity) and one individual-based (e.g., counterfactual fairness). For example, if your system is a loan approval model, you might aim for a demographic parity ratio between 0.8 and 1.2 across protected groups, and ensure that for any individual, flipping their gender attribute does not change the decision more than 5% of the time (counterfactual test).

Practical tools: IBM’s AI Fairness 360 (AIF360) provides 70+ fairness metrics and 10 bias mitigation algorithms. Microsoft’s Fairlearn offers interactive dashboards for visualizing disparities. For production monitoring, Amazon SageMaker Clarify can run scheduled bias detection jobs on inference data. A typical pipeline: during training, run AIF360’s disparate impact analysis on test sets; if the ratio falls below 0.8, apply mitigation like reweighing or prejudice remover. After deployment, schedule monthly Clarify jobs to compare inference distributions against training baselines. Document the chosen thresholds and the rationale—for example, “we selected 0.8 as the threshold because it aligns with the US Equal Employment Opportunity Commission’s four-fifths rule, which is referenced in the AI Act’s recital 44.”

One often-overlooked requirement: Article 10(4) mandates that data sets be examined for possible biases “that are likely to affect the health and safety of persons, have a negative impact on fundamental rights, or lead to discrimination.” This means you must also test for intersectional bias—e.g., how the model performs on Black women versus White men. A 2024 study by the AI Now Institute found that 73% of commercial AI systems tested for intersectional bias failed at least one metric. Use tools like Aequitas (University of Chicago) to generate intersectional bias reports. Store all test results, including failed tests, in your audit trail—regulators want to see that you identified and mitigated, not that you achieved perfect fairness from the start.

5. Designing Human Oversight and Audit Trails That Work in Practice

Article 14 requires that high-risk systems be designed and developed so that natural persons can oversee their functioning. This means implementing human-in-the-loop (HITL), human-on-the-loop (HOTL), or human-in-command (HIC) mechanisms, depending on the risk. For example, a system that recommends medical treatments must have a HITL design where a clinician must confirm each recommendation before it is executed (Article 14(4)(a)). An automated resume screening tool might use HOTL: the system ranks candidates, but a recruiter reviews the top 10% and can override. The Act specifies that the human must have the ability to “decide not to use the system” or “override or reverse the output” at any time.

To comply, you need to log every decision, input, output, and override action. Article 12 requires that logs be kept for at least 6 months, but most experts recommend 2 years to align with GDPR and other EU legislation. The logs must be sufficient to enable traceability of the system’s functioning throughout its lifecycle. For a credit scoring model, each log entry should include: timestamp, model version, input features (anonymised), output score, confidence interval, human reviewer action (approved, overridden, rejected), and the reason for override. Tools like DataRobot MLOps and MLflow Tracking can automate this logging. Ensure logs are immutable—use a write-once, read-many database such as Amazon S3 with object lock or a blockchain-based audit trail (e.g., IBM’s Trust Your Supplier platform offers this for regulated industries).

One practical example: the Dutch government’s SyRI system (benefits fraud detection) was shut down in 2020 partly due to lack of human oversight and auditability. In response, the Netherlands’ Algorithm Register now mandates that all high-risk algorithms publish logs of every decision. Your team should simulate a regulator inspection: pick a random week of logs and verify that you can reconstruct why a specific decision was made, including which version of the model was used and whether a human intervened. If you cannot, your audit trail is insufficient. Budget for at least 100 hours of engineering time to build the logging infrastructure, plus ongoing storage costs (logs can grow 10–50 GB per month for a medium-traffic system).

6. Navigating Conformity Assessment and EU Database Registration

Before placing a high-risk system on the market, you must undergo a conformity assessment (Article 19). For most high-risk systems listed in Annex III, the provider can self-assess against the requirements in Articles 8–15 and 17, then issue an EU declaration of conformity (Article 47).