LLM Driven Processes to Foster Explainable AI

Explainability through "auditable artifacts": Rather than presenting final decisions as black boxes, this system breaks down LLM reasoning into traceable intermediate steps—including factor identification, impact matrices, and game-theoretic analyses—that can be inspected and verified by humans, addressing a key concern in responsible AI deployment.

Moderate alignment with human experts: In a logistics case study, the system achieved 55.5% agreement with human experts on identifying relevant factors and 57% agreement on role classification, suggesting the approach captures expert reasoning but with meaningful gaps; this transparency about limitations is important for responsible implementation.

Combining AI with deterministic analysis: The system pairs LLM components with rule-based mathematical tools (equilibrium analysis, matrix classification) rather than relying on LLMs alone, reducing the risk of hallucination or unexplainable outputs in high-stakes decision support scenarios.

Modular design enables customization and oversight: The "swappable modules" architecture allows organizations to configure the pipeline for their specific needs and replace components, giving human decision-makers greater control over how AI supports their processes—a practical step toward responsible AI governance.

Transparent scoring criteria: Using an explicit eight-criterion rubric to evaluate system performance makes the evaluation process itself auditable, helping stakeholders understand not just what the system recommends, but how its recommendations are being assessed.