Fraud detection systems powered by machine learning (ML) often lack transparency, raising concerns about trustworthiness and interpretability. While Explainable AI (XAI) addresses these issues, many methods rely on correlation rather than causation, potentially overlooking true fraud patterns. This study integrates causal discovery with XAI to propose a novel evaluation framework for fraud detection, validated on synthetic and real-world datasets. Our pipeline combines ML models, SHAP-based explanations, and causal feature selection via CD-NOD. Results show that models trained on causally selected features achieve slightly higher XAI alignment on quantitative metrics compared to correlation-based methods, particularly on synthetic data. However, real-world data challenges such as anonymization has lead to limited causal interpretability. This work lays groundwork for trustworthy AI in high-stakes finance, highlighting the need for dynamic causal methods and improved causality-specific evaluation metrics.