A layered risk scoring and decision support system for customer refund decisions. Agents were approving everything by default because they had no objective information to do otherwise. Built a six-layer deterministic architecture: experience-level anomaly check, policy gate, customer risk profile (six signals, lifetime history with decay weighting), request evaluation, classification, and live case management — with an LLM layer for agent communication. ~88% reduction in agent hours; fraudulent approval rate from ~5% to ~2% target.