Application of AI in corporate finance
Loading...
Date
2026-09-28
Authors
Σολωμός, Χαράλαμπος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This thesis evaluates how reliably three large language models (LLMs), ChatGPT, Claude and Gemini, perform selected corporate finance tasks using FY2025 Form 10-K evidence from Procter & Gamble, Church & Dwight and Edgewell. The controlled design covers four task tiers: financial data extraction, ratio analysis, quantified adverse developments and a fixed-assumption valuation based on free cash flow to the firm (FCFF). Each tier is run with named and partially anonymized inputs, and every model–company–condition–tier combination is repeated three times, giving 216 runs. All models received identical prompts in fresh conversations, with memory and web search disabled, and their responses were scored against predefined benchmarks. The analysis combines task-specific scores, consistency across repetitions and a supplementary review of errors. Overall mean scores are 98.75% for ChatGPT, 97.41% for Claude and 96.69% for Gemini under the study’s rubric. These are descriptive results for the tested configurations, with differences that vary by task. Every ratio-analysis run receives full marks, and the controlled valuation is also highly accurate. The remaining errors mostly concern disclosure scope, financial attribution and input definitions: a correctly copied earnings figure can still answer the wrong accounting question. Performance does not fall steadily across tiers, and removing overt identifiers is not associated with a consistent score difference. The study contributes to AI-assisted financial analysis by connecting benchmark accuracy to the specific verification an analyst needs at each stage, from selecting a disclosed item to converting enterprise value into equity value. It does not measure human analyst performance, productivity or professional readiness. The observed errors identify specific points where apparently correct financial answers require verification. Numerical and semantic fidelity can diverge; in one valuation run, a debt-input discrepancy below 0.5% pushed equity value per share outside the benchmark tolerance. The findings support supervised assistance with clearly specified tasks and traceable evidence.
Description
Keywords
Generative AI, Large language models, Corporate finance, Financial statement analysis, Partial anonymization, Fcff valuation