Persistent and Conversational Multi-Method Explainability for Trustworthy Financial AI

Multiple explanation methods provide better confidence: The paper demonstrates that comparing explanations from three different methods (LIME, occlusion-based scoring, and saliency heatmaps) helps users assess whether an AI's reasoning is robust. This "triangulation" approach means financial decision-makers can have greater confidence when multiple methods agree on why an AI made a particular prediction.

AI explanations need to be searchable and persistent: Rather than generating explanations on-demand, the system stores explanation artifacts with metadata and summaries, allowing financial institutions to retrieve and audit past explanations. This addresses a critical need in regulated industries where institutions must document and justify AI decisions over time.

Conversational AI can reduce hallucinations in explanations: Using constrained prompting techniques, the system reduced hallucination rates by 36% and increased proper citations to explanation methods by 73%. This is significant because financial institutions cannot afford AI systems that confidently provide false or unsupported explanations to decision-makers.

Automated quality checks are essential for trustworthy AI: The paper evaluates explanations using automated checks for completeness, hallucinations, and proper method attribution. This suggests that responsible AI deployment in finance requires built-in verification mechanisms rather than relying solely on human review.

Human-centered design matters for regulated industries: The architecture prioritizes conversational accessibility and natural-language dialogue, recognizing that financial decision-makers need to understand why an AI recommendation was made—not just receive the recommendation itself.