Validating AI-Driven Nudge Recommendations: A/B Testing Two-Tower and Bandit Models in Simulated Digital Banking Environment

While radical moves towards personalization characterize increasingly ambitious digital banking campaigns, recommender systems are typically based on either collaborative filtering or content-based filtering, and they often prove ineffective due to a lack of behavioral awareness, as well as issues such as high sparsity, cold start, and historical signal sparsity. To overcome this gap, this work proposes an AI-based Nudge Recommendation Model that incorporates a Two-Tower Network (TWN) for static personalization along with a Multi-Armed Bandit (MAB) for real-time adaptive nudge optimization. We tested the model using a controlled A/B experiment in a mobile banking simulator with verified bank customers. Results show substantial behavioral improvements in the treatment group using the AI-driven system, including an increase in overall purchase behavior (48.6% to 62.07%) and a more than fourfold rise in recommendation-following rates (13.6% to 52.87%). Statistical testing confirms the significance of these differences (χ² = 6.49, p < 0.0108). These findings validate the causal impact of personalized and adaptive nudging on user decision behavior and provide a reproducible machine learning behavioral economics model with implications for improving digital banking engagement.