This paper presents a hybrid forensic auditing framework that integrates Large Language Models (LLMs), Graph Neural Networks (GNNs), Retrieval-Augmented Generation (RAG), and a rule-based Finite State Machine (FSM) inspired by Automata Theory to detect corporate fraud hidden within unstructured financial data. Traditional fraud detection systems primarily rely on structured data and predefined rules, which often fail to identify narrative inconsistencies and behavioral patterns present in audit trails, financial statements, and corporate communications. The proposed framework utilizes LLMs to identify semantic and linguistic indicators of deception, while the FSM models fraud progression through interpretable state transitions to support audit traceability and explainability. GNNs are employed to analyze relationships among entities and identify collusive activities within organizational networks, including suspicious communication and transaction patterns. In addition, the RAG module retrieves relevant financial regulations, audit precedents, and compliance guidelines to enhance contextual understanding and decision-making. By combining linguistic intelligence, relational reasoning, and rule-based symbolic analysis, the framework provides an explainable and scalable approach to corporate fraud detection. The framework was evaluated using real-world datasets, including U.S. Security and Exchange Commission filings and the Enron email corpus, and achieved an accuracy of 96%, precision of 95%, recall of 94%, F1-score of 94%, and area under the receiver operating characteristic curve of 0.94, demonstrating improved fraud detection performance, auditing transparency, and regulatory compliance.
Patil et al. (Mon,) studied this question.