When AI Actually Touches Money

When you return a damaged monitor to Amazon and the chatbot offers you ten bucks to keep it, is that real AI making a decision? This episode breaks down the difference between a state machine dressed up as a chatbot (your VPN cancellation offer) and an actual LLM in the loop (Amazon's return-less refund system). We walk through the guardrail architecture: constrained output schemas, rule engines that map classifications to pre-set percentages, hard caps per item class, confidence thresholds, fraud-scoring overrides, and human escalation queues. The AI never touches the dollar amount — it classifies damage into tiers, and the old authorization ladder does the rest. Plus: the adversarial attack surface when customers learn to prompt-engineer their refund descriptions. Episode #936293 — open it directly at myweirdprompts.com/936293