Asian Journal of Research in Computer Science

This study develops and validates a conversational AI–powered fraud prevention framework for real-time transaction security in augmented reality (AR) e-commerce environments. Adopting an explanatory sequential mixed-methods methodology, the research integrates architectural framework design, quantitative machine learning evaluation, and qualitative trust analysis to address limitations of transaction-only fraud systems in immersive commerce. Publicly available credit card transaction datasets were augmented with synthetically generated AR conversational logs, enabling the modeling of sentiment, behavioral, and temporal context during checkout interactions. A stacking ensemble architecture combining fine-tuned BERT, XGBoost, and Random Forest classifiers was implemented and evaluated under strict real-time constraints. The proposed framework achieved 98.94% accuracy, a 0.935 F1-score, 0.968 ROC-AUC, and a mean inference latency of 44.8 ms, while correctly approving 99.85% of legitimate transactions. Conversational sentiment features proved highly discriminative, contributing 19.5% of total predictive importance and significantly enhancing contextual risk assessment and user trust. The results demonstrate that conversational AI can simultaneously deliver robust fraud detection, sub-100 ms responsiveness, and frictionless user experience. The framework is immediately deployable in regulated domains such as finance and healthcare, where explainable, low-latency security decisions are critical for compliance and user confidence.