Ontology Framework Boosts Auditable LLM Analytics in Finance

A new framework, the Knowledge-Driven Analytics Framework (KDAF), has been developed to address the critical need for trust and auditability in large language model (LLM) applications within enterprise finance. In regulated environments like Financial Planning and Analysis (FP&A), the usability of an LLM's answer hinges on its traceability to authoritative sources and its post-facto auditability, rather than just its fluency or accuracy. KDAF builds ontology-driven knowledge systems through an iterative six-stage process. KDAF employs Context-Aware Relevance Propagation (CARP) for evidence retrieval, ensuring that every retrieved fact comes with its relationship type, confidence score, and full source lineage. This contrasts with traditional retrieval methods that might prioritize accuracy without providing clear provenance. Evaluations on FinanceBench showed that while retrieval-augmented conditions significantly outperformed zero-context inference in correctness, KDAF's accuracy was statistically similar to simpler lexical retrieval methods like BM25. However, KDAF excelled in auditability, achieving the highest citation traceability F1 score and ensuring that all selected evidence resolved to a complete provenance chain. This research argues that for enterprise finance, auditability is the primary justification for using ontology-grounded retrieval, not just accuracy.