Explainability Assistant: A Conversational XAI Interface for Interpreting Energy Consumption Models

Conversational interfaces dramatically improve XAI accessibility: The Explainability Assistant achieved 94% accuracy in understanding user questions about AI models—a substantial improvement over previous conversational XAI systems like TalkToModel (76.8%)—suggesting that natural language interaction makes AI explanations more accessible to non-technical users like building operators.

Domain experts strongly prefer conversational over traditional dashboards: All energy specialists evaluated in the study unanimously preferred the conversational interface to traditional XAI dashboards for practical use, indicating that how explanations are delivered matters as much as the explanations themselves for real-world adoption.

Modern AI can make complex AI more interpretable: By using Large Language Models' function-calling capabilities, the system can explain predictions from complex "black box" models (like Genetic Programming-based forecasters) without requiring users to have technical expertise, democratizing access to model interpretability.

Flexible, context-aware explanations are more useful than one-size-fits-all dashboards: The conversational system adapts to different questions and contexts without task-specific programming, allowing facility managers to ask follow-up questions and get personalized explanations rather than being limited to pre-designed dashboard visualizations.

This approach generalizes beyond energy forecasting: The system works across different machine learning problem types without custom fine-tuning, suggesting conversational XAI could improve interpretability across many industries where non-technical professionals need to understand AI decisions.