Bolstered by the onward march of digitalisation, advances in artificial intelligence (AI) are delivering new value propositions for businesses. In some industries, including the financial audit industry, AI is emerging as a key competitive technology. And the majority of AI innovations currently influencing many aspects of our lives are based on “deep learning”.
Deep learning is a comparatively youthful subdiscipline that effectively represents the next stage in the evolution of machine-learning systems. Generally speaking, deep learning refers to the use of self-learning (also known as adaptive) artificial neural networks. Many of the basic ideas behind deep learning are not new at all, but can be traced back to the invention of artificial neural networks or neurons in the 1940s and 1950s. Here, the qualifier “deep” describes a learning process based on huge numbers of artificial neurons, often structured in several hundred layers. In contrast to conventional “machine learning”, deep learning is capable of autonomously extracting the attributes of greatest relevance to a given task from a corpus of raw data. This obviates the need for a time-consuming manual extraction process that relies on human expertise. This ability to learn solution models while simultaneously learning or identifying relevant data attributes is known as “end-to-end learning”.
Currently, audit practice is being transformed by the migration from IT-assisted auditing to what is increasingly becoming AI-assisted auditing. The breadth of potential applications for deep learning at various stages in the auditing process is already becoming apparent.
Deep learning methods make it possible to analyse high volumes of structured data (held in e.g. database or Excel tables) to obtain additional audit assurance. In the context of journal entry testing (as described in ISA 240), this makes it possible to distinguish regular from irregular journaling systems. Increasingly, these processes are also being used as a part of analytical audit procedures (ISA 520) to forecast corporate performance indicators. Finally, it is also possible to use deep learning to “acquire” a representative sample from a large population of accounting or business-process data (ISA 530).
Deep-learning methods also deliver added value for auditors involved in analysing unstructured data (such as textual, speech or image data). Indeed, deep learning-based sentiment analysis is already regarded as a valuable method for analysing corporate annual and management reports (ISA 720) as well as inquiries (ISA 500). Deep learning-based object recognition is also used for inventory audits (ISA 501), and it is increasingly feasible that suggestions for audit documentation (ISA 230) could be pre-generated in natural language using deep-learning methods.
Eventually, the symbiotic co-evolution of audit activities and deep-learning processes will cause AI to become an established part of audit practice. Our new Certified Audit Data Scientist course addresses this technological transformation and teaches state-of-the-art AI-assisted audit procedures.
This blog post is an excerpt from the authors’ article on “Deep Learning für die Wirtschaftsprüfung – Eine Darstellung von Theorie, Funktionsweise und Anwendungsmöglichkeiten” (Deep Learning for Auditing – A Presentation of Theory, Methodology and Possible Applications) (in: Zeitschrift für internationale Rechnungslegung (IRZ) issue 7/8, 2021, pp. 349-355, C.H. Beck Vahlen Verlag).
Co-author Anita Gierbl completed her doctoral thesis on “Data Analytics in External Auditing” at the University of St. Gallen (HSG). Since 2019, she has been working in external auditing at PwC Switzerland. In addition to training as a Swiss Certified Public Accountant, she works as a Digital Accelerator at PwC Switzerland. She also works as a research associate at the University of St. Gallen and as a specialist technical assistant for Swiss standards body Swiss GAAP FER.