Agentic Explainable Artificial Intelligence (Agentic XAI) Approach To Explore Better Explanation: A Case Study in Decision Support for Rice Cultivation in Japan

Iterative AI refinement has limits: While an AI agent refined agricultural explanations through 11 rounds, quality peaked at rounds 3-4 before declining. Crop scientists rated the explanations 30-33% better after refinement, but excessive iteration actually made them worse than the original explanation, demonstrating that "more refinement" doesn't always mean "better results."

There's a trade-off between detail and usefulness: Early refinement rounds lacked specific, actionable details (bias problem), but later rounds became too detailed and expensive to implement, with the AI making unfounded economic claims (variance problem). This suggests responsible AI systems need strategic stopping points rather than continuous improvement.

Human experts and AI judges disagree on quality: Crop scientists and AI-based judges scored the explanations differently, with AI judges consistently overrating later rounds. This gap highlights a critical challenge: we cannot rely solely on AI systems to evaluate whether explanations are actually useful to real people.

Accessible explanations require human expertise in the loop: The study combined SHAP (a technical explanation method) with an LLM to make rice cultivation recommendations understandable. The fact that human crop scientists were needed to evaluate quality suggests that translating AI into plain language still requires domain experts to validate accuracy and practicality.

Responsible AI deployment needs clear evaluation criteria: The study measured seven specific criteria (clarity, practicality, credibility, etc.) rather than a single "accuracy" score, revealing that good explanations require balancing multiple competing goals—a lesson applicable beyond agriculture.