AI-Based Framework for Real-Time Anomaly Detection and Governance in Multi-Entity Enterprise Financial Data

Abstract

Financial consolidation within modern multinational enterprises involves aggregating high-throughput transactional ledgers across heterogeneous Enterprise Resource Planning (ERP) landscapes, converting foreign functional currencies, and executing complex intercompany eliminations under strict regulatory frameworks (IFRS 10 and US GAAP Topic 810). Traditional static rule-based validation checks and post-hoc manual sampling mechanisms are inherently limited in identifying complex non-linear anomalies, multi-entity booking shifts, and subtle temporal cutoff manipulations prior to financial close. This research presents a production-ready, hybrid artificial intelligence framework specifically engineered for real-time transactional anomaly detection and continuous regulatory governance in multi-entity financial consolidation pipelines. The proposed architecture integrates a supervised ensemble classification layer (Gradient Boosting and Random Forests) trained on historically verified journal entries with an unsupervised deep autoencoder neural network that flags novel structural outliers via reconstruction error metrics. To overcome the interpretability barriers inherent in deep learning, the engine incorporates Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) integrated with an active Human-in-the-Loop (HITL) auditor feedback loop. Evaluated across an enterprise testbed comprising 10 reporting entities across 5 functional currencies, the proposed hybrid model achieves an overall anomaly detection accuracy of 94.2%, a precision of 93.5%, and reduces false-positive rates to 2.8%. Furthermore, the framework reduces close-cycle manual reconciliation overhead by approximately 60 hours per month per entity, demonstrating its operational efficiency and audit compliance capability.