The use of AI in combating fraud and money laundering is met with considerable skepticism by many specialists. One reason is the inherent difficulty in training an effective AI due to the highly imbalanced nature of datasets, where the vast majority of transactions are legitimate. Additionally, in critical sectors like finance and medicine, where incorrect predictions can have severe consequences, the concept of trustworthy AI is crucial for convincing domain experts to delegate part of their work to AI systems. In this context, explainability tools developed by the scientific community offer new perspectives by unveiling the “black boxes” of AI. These tools provide experts with control mechanisms to evaluate whether AI systems can be trusted. In this work, we first address the challenge of training effective AI in imbalanced settings. We then explore the application of various explainability tools (LIME, SHAP, ANCHORS, PyXAI) on AI models trained for fraud detection. The aim is to identify which of these approaches can help build trust in automated systems with domain experts and whether AI models can learn pre-defined patterned rules or can discover new rules, leading to a better understanding of the decision process. This study highlights the importance of combining effective training techniques with robust explainability tools to enhance the overall trustworthiness of AI in sensitive applications. By examining the different aspects, we aim to contribute to the development of more reliable and transparent AI systems that can effectively assist in fraud detection while retaining the trust of human experts.