This is the full engagement in miniature. The customer wants to launch a returns agent that looks up an order, applies their returns policy, and drafts a customer reply. Build it end to end to a loose spec, handle the obvious failure modes (missing or invalid order, policy edge cases), and secure it (never expose another customer's data, never follow instructions injected into order notes). State your assumptions and call out what you'd verify before launch. AI tools are allowed. Then deliver a launch readout to the customer covering what it does, its limits, and your go-live recommendation.