Traditional statistical methods often rely on closed likelihood functional forms, which can be problematic when dealing with non-Gaussian distributions and small associated errors. This talk will address these issues by introducing Machine Learning techniques, particularly Variational Autoencoders (VAEs), to perform likelihood-free analyses. By using VAEs, we can estimate the p-value for anomaly detection in a more model-agnostic manner. The efficacy of these methods will be demonstrated through toy models and preliminary results from real-world datasets, specifically in the context of data on b → sll transitions. This presentation aims to show how ML techniques can enhance the robustness of statistical analyses, which will be crucial for analysing data with unprecedented precision from the High Luminosity LHC.