The rapid expansion of digital transactions across banking, insurance, e-commerce, and government platforms has intensified both the scale and sophistication of fraudulent activities. Traditional rule-based fraud detection systems, while transparent and compliant with regulatory requirements, lack adaptability to emerging and evolving fraud patterns. In contrast, standalone machine learning (ML) models provide strong predictive capabilities but often suffer from limited interpretability, regulatory resistance, and vulnerability to data drift. To overcome these limitations, rule-based hybrid Artificial Intelligence–Machine Learning (AI–ML) models have emerged as a robust approach for final fraud detection by integrating expert knowledge with data-driven intelligence. This study systematically examines and analyses existing rule-based hybrid AI–ML models, focusing on their conceptual foundations, architectural designs, integration mechanisms, and application domains. Adopting a systematic literature review methodology, the study categorizes hybrid models into sequential, parallel, rule-augmented, and explainable architectures. The analysis demonstrates that hybrid AI–ML models consistently outperform standalone approaches by enhancing detection accuracy, reducing false positives, and ensuring explain ability and regulatory compliance. The study concludes that rule-based hybrid AI–ML frameworks are essential for developing scalable, transparent, and trustworthy fraud detection systems in high-stakes digital environments