A Governance-Oriented Evaluation of Explainable AI in Business Analytics

Artificial intelligence (AI) is increasingly embedded in analytics-driven decision systems operating in high-stakes environments, where predictive performance alone may not ensure robustness, traceability, or readiness for governance. While explainable artificial intelligence (XAI) is commonly treated as an interpretability tool, its role as a measurable diagnostic component remains underexplored. This study evaluates explainability beyond predictive accuracy using a controlled dual-pipeline design with identical data, model, and validation settings. Using the UCI Default of Credit Card Clients dataset, an XGBoost model achieved strong predictive performance (AUC ≈ 0.78; AP ≈ 0.56) while maintaining high decision stability under retraining (agreement ≈ 99.49%). Global explanations were highly reproducible across runs (Spearman ρ ≈ 0.99), and entropy-based local explanation analysis revealed substantially higher attribution dispersion in false-negative cases (odds ratio ≈ 6.58), linking explanatory uncertainty to misclassification-prone regions. The findings demonstrate that explainability diagnostics can reveal measurable patterns of stability, reproducibility, and uncertainty that are not captured by predictive metrics alone. The study advances a governance-oriented perspective in which explainability serves as a reproducible diagnostic layer that supports monitoring, validation, and audit-oriented assessment in analytics-driven decision systems.