AI-ENABLED FINANCIAL ACCURACY MODELS FOR IMPROVING ERROR DETECTION AND REPORTING INTEGRITY IN CORPORATE ACCOUNTING SYSTEMS

Abstract

This quantitative study examined the effectiveness of AI-enabled financial accuracy models in improving error detection and reporting integrity within corporate accounting systems. Using a retrospective observational design, the analysis was based on 48,620 accounting transactions extracted from integrated accounting information systems, including general ledger entries, subledger postings, and period-end adjustments. Confirmed error outcomes were identified through documented reversals, reclassifications, reconciliation resolutions, and audit adjustment linkages, yielding an observed error rate of 2.5% (1,215 transactions). Transaction-level and account-period indicators capturing transaction magnitude, posting lateness, exception frequency, reconciliation variance, account-linkage intensity, and workflow complexity were operationalized and analyzed using statistical and machine learning techniques. Descriptive findings showed that error-labeled transactions exhibited substantially higher reconciliation variance (mean USD 1,420 vs 180), longer posting delays (mean 5.6 days vs 1.8 days), and extended correction cycle times (mean 6.9 days vs 1.1 days) compared with non-error transactions. Correlation analysis demonstrated coherent positive associations among timing disruption, exception frequency, and reconciliation variance, while reporting integrity indicators showed strong alignment with audit adjustment volume (r = 0.63). Regression results indicated that posting lateness (odds ratio 1.48), exception frequency (1.36), reconciliation variance (1.29), and account-linkage intensity (1.21) were significant predictors of confirmed error occurrence after controlling for organizational and system factors. Comparative model evaluation showed that AI-based classifiers outperformed baseline rule-based exception detection, with a tree-based ensemble model achieving an AUC of 0.87, precision of 0.49, and F1-score of 0.60, compared with an AUC of 0.68 and precision of 0.22 for the rule-based approach. Overall, the findings demonstrated that AI-enabled financial accuracy models provided superior discriminatory power and more balanced detection performance by capturing multivariate process, control, and linkage-driven risk patterns that traditional rule-based mechanisms failed to identify consistently.