Abstract
The volume and complexity of digital accounting records has made audit planning more and more reliant on analytically defensible methods for prioritizing risky engagements and transactions. This work proposes a leakage-aware, explainable machine-learning framework for intelligent audit-risk assessment and investigates the degree of possible harmonization of predictive discrimination, probability quality, explainability and operational threshold choice into one decision-support architecture. An audit observation dataset was developed and is reproducible based on nine theoretically grounded audit risk indicators: financial discrepancies, prior misstatements, internal-control weaknesses, unusual journal entry activity, abnormal revenue growth, liquidity stress, leverage, auditor change and management override, resulting in 4,000 audit observations. Accuracy, precision, recall, F1-score, ROC-AUC, PR-AUC, Brier score, and cross-validation were used to compare Logistic regression, random forest, Gradient boosting and XGBoost. XGBoost outperformed the other models in terms of the highest ROC-AUC score (0.815), while logistic regression recorded the best PR-AUC score (0.605) and maximum recall at the standard 0.5 threshold (0.691). Gradient boosting obtained the best overall accuracy (0.835) and Brier score (0.123). SHAP analysis showed that the most important drivers of the XGBoost prediction were discrepancy ratio and internal-control weaknesses, globally. The recall of flagged engagements was shown to improve from 35.3% to 56.0% when using a 0.30 threshold instead of the standard 0.50.Threshold analysis revealed that the recall of flagged engagements improved from 11.3% to 22.6% by using a threshold of 0.30 instead of the standard 0.50. The results illustrate that audit-oriented AI needs to be assessed as a socio-technical system that needs to be governed, not only on the basis of its accuracy per se. For high-stakes assurance settings, a defensible implementation requires leakage control, multiple performance metrics, probability calibration, transparent explanations, cost-sensitive thresholds, and explicit retention of auditor judgment.