This study investigates the effect of fly ash substitution on compressive strength in cemented paste backfill (CPB) systems using an integrated framework combining predictive machine learning, explainable artificial intelligence, and causal machine learning approaches. A dataset comprising 116 experimental CPB mixtures prepared with mineral processing tailings from a Pb–Zn underground mine and fly ashes from different thermal power plants was analyzed. The maximum compressive strength reached 10.33 MPa at 90 days of curing. The cross-validated XGBoost model achieved an R² of 0.58 and successfully predicted strength values up to 9.73 MPa. Causal analysis indicated an average treatment effect of approximately 0.28 MPa per 1% fly ash substitution, although the effect showed substantial heterogeneity across physicochemical conditions. Global and local SHapley Additive exPlanations (SHAP) analyses identified curing time as the dominant factor controlling strength development, emphasizing the importance of late-age performance in fly ash–containing systems. The fly ash ratio itself was not a primary explanatory variable; instead, chemical composition, particle size distribution, and specific gravity played decisive roles. These findings demonstrate that fly ash performance in CPB systems cannot be reliably evaluated using dosage-based approaches alone and should be optimized by considering physicochemical characteristics and curing conditions. The proposed explainable and causal data-driven framework provides a practical decision-support tool for sustainable utilization of mineral processing by-products in cement-based systems.