Human-Centered LLM-Agent System for Detecting Anomalous Digital Asset Transactions

Human-centered design improves trust in AI decision-making: HCLA demonstrates that allowing non-experts to express their analytical intent in natural language and inspect the reasoning process step-by-step substantially improves interpretability and decision transparency compared to traditional black-box models, suggesting that involving humans in the AI workflow—rather than just explaining outputs—builds confidence in high-stakes decisions.

Separating evidence from justification enhances accountability: The system's explicit separation of evidence scoring (what the data shows) from expert-style justification (how experts interpret that data) creates a clearer audit trail, which is critical for regulatory compliance and financial forensics where decision-makers need to understand not just what the AI flagged, but why it matters in context.

Explainability alone is insufficient for responsible AI in finance: The paper emphasizes that HCLA goes "beyond explainability" to focus on accountability and regulatory judgment, suggesting that in sensitive domains like financial crime detection, simply explaining a model's predictions is not enough—systems must align algorithmic evidence with real-world investigative and compliance standards.

Open-source, transparent tools support responsible deployment: By implementing HCLA as an open-source, web-based interface rather than a proprietary black box, the authors demonstrate a practical commitment to transparency and auditability, making it easier for regulators and institutions to scrutinize how AI systems make consequential decisions about financial transactions.

Classical methods with human oversight may outperform pure AI automation: The finding that combining traditional anomaly detectors with human-centered reasoning produces better outcomes than the detector alone suggests that responsible AI in high-stakes domains may require hybrid approaches rather than fully automated decision-making.