This research examines how explainable machine learning (XAI) techniques can identify and measure churn risk factors in India’s quick-commerce delivery platforms. Unlike prior studies focused on prediction accuracy, this work emphasises interpretability to ease practical business interventions. Utilizing Blinkit’s operational data, which shows a 47.35% churn rate, the study employs an ensemble method integrating Random Forests, Gradient Boosting, and SHAP analysis to break down churn drivers. The method introduces the Quick-Commerce Customer Stability Index (QCCSI), a composite measure integrating delivery reliability metrics (35%), support interaction quality (25%), engagement frequency (20%), and account tenure dynamics (20%). Analysis of 15,000+ customer transaction sequences reveal that delivery consistency (coefficient:
-0.68, p<0.001) and unresolved support queries (hazard ratio: 2.14) are the strongest predictors of churn. Critically, this study finds that customer dissatisfaction appears not from single failure events but from accumulated friction points. A three-point increase in QCCSI correlates with a 31.7% reduction in 90-day churn probability. Implementation of the proposed early warning system could preserve ₹142 million in annualized revenue for mid-market operators with 100,000 active users.