From Alerts to Audit Evidence: LLM-Augmented Data Quality Governance in Financial Pipelines

Abstract

High-throughput data streams are used to process transactions, detect fraud, resolve reconciliations, manage customer relationships and generate compliance reports in the financial sector. Traditional monitoring approaches primarily rely on rule-based alerts and thresholds to detect anomalies without providing explanations of potential impact and regulatory significance. This can slow down investigations and overwhelm analysts. To overcome this problem, this paper introduces a new data quality governance framework enhanced with Large Language Model (LLM) mechanism that integrates anomaly detection, context-aware reasoning, confidence-based prioritization, schema drift intelligence, and governance reporting. The architecture combines transaction monitoring systems with a Large Language Model (LLM)-based reasoning engine to provide actionable insights on the basis of alerts. Anomalies are augmented with metadata lineage, historical patterns, structural relationships, and criticality to predict impact, determine the cause, and suggest fixes. A schema drift analyzer assesses structural drift and its impact, and an automated reporting engine produces human-friendly reports for auditors and governance. In experimental testing using the Credit Card Fraud Detection, PaySim data sets and synthetic schema use-cases, metadata and system logs, the proposed approach yielded 96.8% precision, 95.1% recall and 95.9% F1-score, with reports 10% better (94%) than previous methods (84%). The findings demonstrate that the proposed approach establishes a connection between technical and business anomalies and offers a practical approach towards next-generation trustworthy financial intelligence systems.