Semantic Integrity Auditor: LLM-Based Semantic Drift Detection for Enterprise KPI Governance

Enterprise organizations depend on dashboards and reports to drive critical business decisions, yet the logic behind those dashboards silently drifts over time. Columns are renamed, filters are added, aggregation methods change, and the number labeled “Gross Margin” stops meaning what the business intended. Traditional data quality tools detect schema violations and null values but cannot answer a more fundamental question: does this metric still reflect its original business definition? This paper introduces the Semantic Integrity Auditor (SIA), a system that uses large language model (LLM) reasoning to compare KPI implementations against a structured registry of business intent. SIA performs dual-layer auditing across analytics surfaces (dashboard formulas, filters, visual configurations) and data pipeline structures (DDL, joins, grain, freshness). Validation across 34 enterprise-realistic scenarios yielded a 96.7% accuracy rate on standard cases, with a total cost of $0.23. The system also demonstrated emergent contextual reasoning, identifying a governance gap that was not explicitly encoded in the registry. These findings suggest that LLM-based semantic comparison represents a viable and cost-effective approach to a class of data governance problems that existing tooling does not address.