Machine Learning-Enabled Continuous Auditing: A Critical Narrative Review

Abstract

Continuous auditing provides ongoing assurance on transaction populations, and machine learning holds the promise of making this practical by detecting anomalies that could not be anticipated in advance. This paper provides a critical review of literature on machine learning-enabled continuous auditing. It is a narrative interpretation of the published research on system architectures, learning techniques, audit assertions, field experience and barriers encountered, and its claims rest on interpretations of the literature rather than numbers. There are four observations to organize the argument. Almost all of the system architectures examined share a common three-tiered organization into acquisition, analytics and workflow, and the learning technique used resides in the analytics tier, while investigation capability resides in the workflow tier. Evaluation of the literature relies on detection and process measures, and very little of the literature reports evidence that such systems affect misstatement detection, audit quality or audit effort. Label provenance, concept drift and the lack of sensitivity estimates cast doubt on the validity of what performance evidence there is, and the most consistent barrier reported is organizational and not technical. This paper proposes a maturity model that would distinguish between claims that could and could not be supported by a system, and a research agenda that would emphasize outcome measure, sensitivity estimates and drift-aware monitoring over more system designs.