video1.0<iframe src="https://www.loom.com/embed/5b819e735c6647efb1411b7e55e3a6c3" frameborder="0" width="1728" height="1296" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>12961728Loomhttps://www.loom.com12961728https://cdn.loom.com/sessions/thumbnails/5b819e735c6647efb1411b7e55e3a6c3-60fd7efd935d440a.gif295.635This Loom explains how RefundPilot, an open source local AI agent, automates customer refund requests using a strict policy and an audit backed admin dashboard. It applies four rules: three or more requests in 30 days flags fraud, final sale items never qualify, requests outside 30 days are denied, and orders over $500 are escalated to a human. The demo shows scenarios with decisions including 2 escalated, 2 denied, 1 approved, and 2 restated, with full execution traces in Langsmith and run metrics for tokens, latency, and costs. A run finished in 3.47 seconds, with 96 percent spent generating the response, and 4.5k tokens served from cache mini, reducing new input to 534 tokens and output to 300 tokens.