VERITAS: A Verified-Data Machine Learning Approach to Segment-Specific Tax Audit Planning

Innovations in artificial intelligence are reshaping how tax administrations approach compliance and audit planning, yet existing AI-based fraud detection studies largely treat the taxpayer population as homogeneous or remain conceptual frameworks awaiting empirical validation. This gap is consequential because audit resources are limited, evasion tactics are increasingly sophisticated, and misallocating scarce audit capacity carries a direct fiscal cost. To address it, this study presents VERITAS, a machine learning-based decision support system operationalizing a segment- and sector-aware architecture for corporate income tax audit planning: a single-layer model for Large Taxpayer case selection, and a novel two-layered model for small and medium enterprises (SMEs) that filters evasion-suspect cases before prioritizing them by expected tax-recovery yield against a target threshold. Ten classification algorithms were compared across 4063 SME and 1903 Large Taxpayer financial statements, with correlation-ranked feature selection subsequently applied to each. Random Forest consistently outperformed all alternatives across every segment, sector, and task examined; feature selection improved performance in every case; sector-specific modeling outperformed a generic classifier in two of four SME sectors, matched it in a third, and was marginally outperformed in the fourth; and a novel business-activity-code feature was retained in most analyses. These findings position VERITAS as a practical innovation in tax audit practice: an architecture for AI-driven audit planning that is internally validated against verified audit outcomes within the historical Lebanese dataset examined, built entirely from data tax administrations already collect, and potentially transferable to comparable jurisdictions, though not yet operationally deployed.