Authors: Dr.S.Narmatha

Abstract: Manual effort continues to be involved in the process of analyzing the financial statements through the process of reconciling the structured XBRL format to unstructured text-based disclosure. This paper describes a novel multi-agent framework for AI-assisted financial report analysis involving the automatic extraction, validation, and synthesis of financial reports. The framework uses three different agents: a parser agent for XBRL and PDF extraction; a reconciliation agent for cross-statement validation; and a narrative agent using a retrieval augmented generation approach to analyze disclosures. The framework is benchmarked against manual workflows of analysts through 10 fiscal filings using the criteria of detecting inconsistency, completeness of fields extracted and consistency of narratives with the original document. It is shown that the specialization of agents leads to fewer cross-statement inconsistencies than those found through baseline single-model methods, and constraining the narrative agent with a retrieval component enhances factual consistency of generated narratives with original sources.

DOI: https://doi.org/10.5281/zenodo.24246365