Beyond the bottom line: using generative AI to extract narrative financial disclosures

While traditional machine-readable datasets (such as Compustat) have advanced financial capital markets research, a significant amount of valuable data remains locked in unstructured, PDF-based financial disclosures. This limitation hampers the scalability of archival research. Earlier efforts using keyword-based heuristics and traditional statistical and machine learning models have faced difficulties in addressing the variability and complexity of financial reporting formats. This study proposes an AI-based scalable framework for extracting structured data from unstructured financial reports. Leveraging a context-aware prompting strategy with large language models (LLMs), the framework identifies relevant page ranges, isolates targeted disclosure sections, and extracts structured quantitative information. To demonstrate its practical utility, we apply the framework to retrieve capitalization thresholds disclosed in the “Capital Assets” section of the note disclosures within governmental Annual Comprehensive Financial Reports (ACFRs). An evaluation on a sample of county and municipal ACFRs yields 99% accuracy in identifying relevant page ranges and 94% accuracy in extracting data at the individual note level. In addition, we replicate the methodology to extract Valuation Allowance for Deferred Tax Assets (DTA) from 10-Ks with 100% success. Our methodology contributes to (1) the accounting literature by enabling the construction of novel datasets for large-sample archival research, (2) the Accounting Information Systems (AIS) literature by demonstrating how emerging technologies can address complex data extraction challenges, and (3) practice by enhancing the efficiency of regulatory oversight and public sector monitoring.