video1.0<iframe src="https://www.loom.com/embed/b7a9250b273e4b21964d9adfc757d870" frameborder="0" width="1728" height="1296" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>12961728Loomhttps://www.loom.com12961728https://cdn.loom.com/sessions/thumbnails/b7a9250b273e4b21964d9adfc757d870-59f1f932b6ab8088.gif559.151This Loom explains how an AI refund agent can approve or deny e-commerce refunds using a deterministic policy engine that cannot be overridden by customer pressure or prompt injection. It demonstrates an approve case for an unopened return within 30 days under policy 2.3, then a denial for a Jordan account flagged for abuse under 2.6, and shows legal threats do not change the result because the system recalculates the refund amount from order data. It blocks direct prompt injection attempts before the model runs, routes refunds over $500 to human sign off, and escalates when confidence drops below a threshold. The author also highlights an evaluation harness with 23 scenarios, including 8 adversarial attacks, achieving 100% decision accuracy and 100% injection detection.