Deep learning meets risk-based auditing: A holistic framework for leveraging foundation and task-specific models in audit procedures

Abstract

• How does the manuscript address issues in accounting and information systems? The manuscript delves into challenges in the realm of accounting and information systems with a particular focus on the application and implementation of Deep Learning (DL) into auditing procedures. Consequently, our work can be situated within the interdisciplinary field of research that bridges cutting-edge technologies like DL with the traditional field of auditing. We have developed a framework, providing comprehensive application and implementation guidelines for the effective deployment of DL by auditors. • What are the theoretical underpinnings and why is the theory appropriate for the research question? The development and evaluation of our DL Framework is rigorously based on Design Science Research (DSR). More concretely, we developed the DL Framework based on the theoretical foundations of the CRISP-DM Process Model and the initial phases of a risk-based audit derived from the International Standards on Auditing (ISAs). Moreover, to incorporate a DL perspective throughout our framework, we have carefully tailored each phase of the developed DL Framework toward the functionalities of DL in alignment with the ISAs. Furthermore, we conducted focus group discussions and expert interviews to empirically extract some DL requirements defining how to implement this complex DL technology within organizations. For the extraction of DL requirements, we utilized an innovative approach for qualitative data analysis incorporating a Large Language Model to analyze the transcribed qualitative data. We evaluated the DL Framework with subject matter experts based on established DSR evaluation criteria. • Why are the research methods used appropriate for the research question being asked? Our rigorous development and evaluation process is based on DSR, which empowers us to address important practical issues through the creation of our central artifact. • What makes the research novel, interesting, and defensible? This is the first rigorously developed and evaluated DL Framework for the auditing profession outlining overarching application and implementation guidelines to augment existing risk-based auditing. • What are the implications for future research and practice? The developed DL Framework can be further evaluated by case studies adopting and integrating the framework in practice. Deep Learning (DL) is a technology with potential to enhance effectiveness and efficiency in audit procedures. However, due to the complexity of DL, it still lacks widespread adoption in the auditing profession. To address this problem, this study follows Design Science Research (DSR) methodology to develop a DL Framework. It integrates the phases of the Cross Industry Standard Process for Data Mining and the phases of risk-based auditing, giving particular attention to guidance for DL. The DL Framework developed here provides detailed application advice on how to augment risk-based auditing with DL. Facing challenges in the realization of our developed DL Framework, we obtained feedback from auditors and computer scientists and identified four key implementation requirements for our DL Framework within organizations. Additionally, we assessed the overall usefulness of our DL Framework based on established DSR evaluation criteria.