RESHAPE: Explaining Accounting Anomalies in Financial Statement Audits by enhancing SHapley Additive exPlanations

AI transparency is critical for financial audits: The paper demonstrates that while deep learning models can effectively detect accounting anomalies, auditors cannot adopt these tools without understanding why the model flagged something as suspicious. This requirement for explainability is not optional but essential for real-world deployment in regulated industries.

Current explanation methods fall short for auditors: Existing explainability techniques like SHAP can explain what individual encoded variables contributed to a model's decision, but this level of detail is too technical for human auditors to act on. The paper shows that explanations need to be translated into business-relevant terms (attributes) that auditors can actually use.

RESHAPE bridges the gap between AI and human decision-making: The proposed method aggregates technical explanations into attribute-level insights—essentially translating "machine language" into "auditor language." This represents a necessary step toward making unsupervised deep learning tools practically usable in financial auditing.

Responsible AI requires domain-specific solutions: The paper illustrates that generic explainability methods are insufficient; financial auditing has unique requirements that demand tailored approaches. This suggests that responsible AI implementation across industries may require customized transparency solutions rather than one-size-fits-all tools.

Evaluation frameworks matter for responsible deployment: The authors introduce a framework to compare different XAI methods' effectiveness in auditing contexts, emphasizing that responsible AI adoption requires not just better explanations, but systematic ways to measure whether those explanations actually serve their intended purpose.