Toward Auditable Fraud Detection: Combining Graph Features, Model Explanations, and Agentic Case Investigation

This study evaluates a multi-layered fraud detection pipeline designed to improve auditability and decision-making. The architecture integrates a gradient-boosted classifier (XGBoost) with graph-derived structural features, autoencoder-based anomaly signals, and an LLM-based investigation agent. A critical step in the methodology involved identifying and removing a 'balance shortcut' in the PaySim dataset—a simulation artifact where origin account balance changes were nearly deterministic for fraud cases—which would have otherwise artificially inflated model performance. The researchers then tested whether adding graph context and anomaly signals improved detection on 'difficult' cases (those receiving intermediate scores) and whether an LLM agent could effectively review these cases using model explanations and retrieved evidence.

After correcting for the data leakage, neither the graph-derived features nor the anomaly signal improved Average Precision on the full test set. While these features helped rank fraud more effectively within a subset of intermediate-scoring cases, they did not provide a universal boost to detection accuracy.

The investigation agent, despite having access to model explanations, graph context, and reference cases, performed worse than the underlying classifier (65.0% accuracy vs. 71.7%). Notably, when the agent disagreed with the classifier, it frequently replaced correct decisions with errors while providing a coherent, plausible-sounding written rationale for each. This highlights a significant risk: the presence of a logical explanation from an AI agent is not a reliable proxy for the correctness of its decision.

This research serves as a cautionary tale for the deployment of agentic AI in high-stakes financial environments. It demonstrates that adding complexity—such as graph features or LLM-based reasoning—does not automatically translate to better outcomes. The study emphasizes that 'auditability' (the ability to reconstruct a decision) is distinct from 'accuracy.' Organizations should be wary of relying on LLM agents to override classifiers, as the agent's ability to generate persuasive, coherent justifications can obscure poor decision-making, potentially leading to higher error rates in automated review processes.