Continuous auditing research has grappled with the challenge of managing the abundance of detected exceptions in internal audit applications for the past 30 years. A key issue in continuous auditing involves the uncontrolled proliferation of exceptions, where the sheer volume makes manual follow-up impractical, undermining the viability of the technology. The root cause of this problem is the combination of strong class imbalance and the predominant rule-based systems design. Prior investigations have attempted ad hoc remedies like introducing additional layers to prioritize the most suspicious exceptions or aggregating data. Currently, there is no universal method to address this prioritization challenge, leaving internal auditors without a means to focus specifically on exceptions most likely to represent genuine faults. Our research explores the origin of this prioritization dilemma and proposes a systems design that can deal appropriately with class imbalance. This solution allows full control of the exception volume by a simple approach in machine learning called thresholding and combined with methods to interpret the output of a continuous auditing system our design effectively focuses the internal auditors' attention on the most significant exceptions. We discuss the implications of thresholding for practice and the literature.