Integrating Transparent Models, LLMs, and Practitioner-in-the-Loop: A Case of Nonprofit Program Evaluation

Transparency builds trust in nonprofit AI adoption: The paper demonstrates that public and nonprofit organizations are more willing to use AI tools when models are interpretable rather than "black boxes." By using decision trees—models that show their reasoning—alongside LLMs, the researchers created a system that organizations could understand and trust enough to actually implement.

Practitioners must be involved at every stage: The study emphasizes that AI systems work better for real-world nonprofit work when the people who actually do the work (program staff, evaluators) participate in feature selection, model design, and explanation review. This "practitioner-in-the-loop" approach ensured the AI provided guidance that was genuinely useful for case-by-case program decisions, not just aggregate statistics.

Combining different AI approaches improves both accuracy and usability: Rather than relying on a single AI method, pairing transparent decision trees with LLMs achieved better predictive accuracy while maintaining interpretability. The decision trees provided clear logic, and the LLMs translated that logic into practical, case-level insights that nonprofit staff could act on.

Case-level guidance matters more than aggregate patterns: Traditional AI analysis often shows overall patterns but doesn't help with individual decisions. This paper shows that nonprofits need actionable insights for specific program participants—information that can directly inform whether a student needs additional support, for example—making the hybrid approach more valuable for real program evaluation.