You're deployed at a customer whose refund agent tries to do everything in a single LLM call: look up the order, check it against refund policy, and draft a response. It's unreliable, sometimes skipping the policy check or inventing order details. Rebuild it as a small, explicit multi-step workflow where each step (retrieve order, check policy, draft reply) is verified before the next runs, so it can't fabricate an order or skip a check. State your assumptions. AI tools are allowed. Then scope the phased rollout with the customer's project manager.