In today’s competitive business environment, maintaining customers is a top priority, particularly in the Order-to-Cash (O2C) process, when losing even one customer may have significant financial consequences. This paper presents a rule-based machine learning approach to predict multi-class customer churn, utilizing transactional data sourced from an SAP ERP system. We develop a systematic strategy for identifying consumers who are most likely to walk away by classifying them as Active (0), At-Risk (1), or Churned (2). The effectiveness of machine learning models in churn prediction is demonstrated by a comprehensive literature analysis, which also emphasizes the significance of behavioral analytics and customer segmentation in predictive modeling. This study includes extensive feature engineering techniques and data pre-processing to find key indications, such as delivery trends, purchase patterns, and payment clearing duration. To determine the best accurate prediction model, a thorough assessment of a few machine learning classifiers was carried out, including Random Forest, Gradient Boosting, Extra Trees, CatBoost, and LightGBM. Finally, Gradient Boosting model emerged as the most effective, achieving a macro-averaged accuracy of 95.2%, an F1-Score of 0.954, and a PR-AUC of 0.997. The results provide valuable insights for businesses looking at establishing targeted retention strategies. Businesses may proactively identify customers who could be at risk of attrition, improve their engagement tactics, and eventually increase long-term profitability by integrating predictive analytics into many ERP systems. This study enhances the field by presenting a strong, rule-based churn prediction model designed for O2C operations. The findings emphasize the growing importance of AI-powered decision support systems in customer relationship management, which offer a scalable solution for minimizing churn while preserving revenue growth.