Artificial intelligence in auditing financial results: methodological aspects and practical implementation

Abstract

The article addresses the development of methodological foundations for applying artificial intelligence in auditing financial statements. The necessity of implementing machine learning technologies and large language models to process growing volumes of textual disclosures in financial statement notes is substantiated. The research identifies systematic challenges of modern audit information environment, including exponential growth of textual disclosures, heterogeneity of data sources, complexity of multi-level judgments, new regulatory requirements for digital reporting, and time constraints of audit engagements. Traditional analytical procedures developed in the era of predominantly paper-based document flow reveal systematic limitations in their ability to provide appropriate level of audit assurance while maintaining economic feasibility. The LLM-AuditBridge methodology for semantic analysis of notes in accordance with international auditing standards is proposed. This methodology implements a multi-stage information processing workflow where each stage corresponds to specific audit procedures and generates specialized artifacts that comply with documentation requirements. The architecture includes three key blocks: risk identification analysis, analytical procedures for cross-validation of textual and numerical data, and generation of responses to identified risks. The methodology employs sophisticated prompt engineering principles, including role establishment, context provision, task specification, output format definition, few-shot learning examples, constraint setting, and chain-of-thought reasoning activation. Critical three-level validation system ensures compliance with requirements for sufficiency and appropriateness of audit evidence through cross-prompt validation, documentary validation, and logical consistency checking. Classical machine learning methods, particularly Benford's Law analysis and logistic regression modeling, are integrated to complement LLM capabilities in numerical data analysis. The architecture of a hybrid system combining LLM analysis of textual data with classical statistical methods for numerical indicators is developed. The integration module performs weighted aggregation of scores from three independent sources and convergence verification, where alignment of multiple independent signals dramatically increases confidence in actual issues. Practical recommendations for implementation of proposed approaches into audit practice are presented, including technical considerations for tokenization, API response times, cost analysis, and organizational aspects. The research demonstrates that AI integration in audit procedures is not only technically feasible but also methodologically justified according to international auditing standards, while preserving auditor's professional judgment and skepticism as fundamental principles.