<![CDATA[The intersection of Large Language Models (LLMs) and financial report auditing has rapidly evolved into a substantive area of peer-reviewed academic inquiry and industry experimentation. Grounded in the theoretical lenses of Agency Theory, Audit Risk Theory, and Socio-Technical Systems Theory, this review synthesizes 20 peer-reviewed sources (2023–2026) across five thematic streams: automated auditing pipelines, regulatory compliance verification, fraud detection, LLM benchmarking, and practitioner perceptions. The central conceptual contribution of this review is the positioning of LLMs not as autonomous auditing agents but as probabilistic cognitive augmentation tools — systems that extend human auditors' analytical reach while operating under mandatory human accountability. We find that while current LLMs demonstrate meaningful capability in error detection, compliance matching, and fraud screening, they consistently fall short in domain-specific accounting reasoning, explainability, and regulatory standard citation. The persistent challenge of hallucination, the absence of auditable reasoning chains, and the structural equity gap between large and small audit firms collectively represent the primary barriers to professional-grade LLM deployment. Future research priorities include domain-adapted models, PRISMA-calibrated benchmarking, and harmonized international AI governance for auditing.]]>
Copyrights © 2026
