The limits of autonomy of AI systems in accounting

The aim of this article is to define the acceptable boundaries for delegating decision-making competencies to AI-based systems in the field of accounting and to assess the degree of consistency and objectivity with which advanced language models carry out accounting and auditing processes. Literature analysis and experimental simulations enabled the identification of key factors influencing the discrepancies in assessments and conclusions generated by these systems. The research results revealed that although artificial intelligence (AI) tools can accelerate the processing and interpretation of financial data, their operation is characterized by significant variability in outcomes. The models demonstrated a strong dependence on the assumptions made, the questions posed, and the auditor’s interpretations. From a legal and ethical standpoint, delegating decision-making responsibilities to AI systems encounters substantial limitations due to the lack of legal personality of AI systems and their inability to apply critical expert judgment. Consequently, despite potential operational benefits, these tools should be employed solely as auxiliary support for accounting experts and not as standalone decision-making systems. The main limitations of this study were technical difficulties in analysing large-volume documents and the limited scope of simulations. The conducted research forms the basis for further analyses aimed at optimizing the integration of artificial intelligence in the financial environment while ensuring compliance with legal and ethical norms.