The rapid adoption of Unified Payments Interface (UPI) systems in India has led to a significant rise in digital financial transactions, accompanied by an increase in fraudulent activities. This research proposes an AI-based fraud detection framework using the XGBoost machine learning model to identify fraudulent UPI transactions effectively. Transactional features such as amount, frequency time patterns, and device-related attributes are analyzed to capture abnormal behavior. XGBoost is selected for its ability to handle imbalanced datasets, non-linear relationships, and high-dimensional data. Experimental results demonstrate improved detection accuracy, recall, and ROC-AUC compared to traditional models, highlighting the suitability of XGBoost for real-time UPI fraud detection systems.