After years of discussions and negotiations, in December 2023, the European Council and the European Parliament agreed on shaping the first rules in the work to govern AI- aka, the AI Act. In the coming months, the proposed draft, which adopts a risk-based approach to AI governance, is set to become a law enforced across the EU member states, with far-reaching global effects. In tangible terms, violations would cost companies between €7.5-€35 million or 1.5%-7% of their global turnover, based on the risk level of the AI applications. The current text states that high-risk AI systems must undergo a conformity assessment based on extensive technical documentation before being used in production. The contents of the required technical documentation are defined at a very high level and will need to cover aspects such as accuracy, robustness, and human oversight of AI systems. This talk will analyse the requirements of an AI system’s technical documentation from a data scientist’s perspective. In particular, we will first tackle the technical aspects that must be verified to consider an AI system as conforming. For example, we will discuss human oversight from a technical point of view, highlighting different interpretability techniques that can be used to perform this analysis. We will then show a practical example of technical documentation related to an illustrative machine learning model from a financial use case. This use case will demonstrate how to translate most of the requirements included in the technical documentation into machine learning model tests implemented using open-source Python libraries. By the end of the talk, participants will have a practical understanding of what the new regulation means from the perspective of a data scientist developing new machine learning-based systems.