Balancing the Continuous Auditing Efficiency-Effectiveness trade-off

The proposed project investigates how data aggregation and controlling the threshold level of machine learning algorithms impact the performance of continuous auditing (CA) systems based on artificial intelligence/machine learning. We investigate the extent that the volume of detected exceptions can be controlled with the two design features.

Controlling the volume is regarded as one of the main challenges of previous CA systems typically developed with rule-based classifiers and static software because this lack of control effectively incapacitates CA.

We develop CA systems in three large Swedish organizations and conduct machine learning experiments in which we vary the data aggregation level of real accounting data and threshold levels (thresholding) for nine different machine learning algorithms.We are an interdisciplinary group (auditing + computer science), 90+ publ., cooperating with two professors in computer science with 260+ publ. in comp. sc./artificial intelligence.Contribution/importance: (1) Our investigation of the impact of data aggregation and thresholding on CA performance is unique and may offer a generally applicable solution to the exception detection volume problem with CA systems. (2) This study contributes by a unique investigation of internal auditors´ ability to interpret CA systems´ inner logic affects how they form an opinion from systems’ output, using real CA systems and state-of-the-art machine learning interpretability models.