I built a WhatsApp AI agent for a local clothing shop in Quetta. Customers text it in Roman Urdu like they would a shopkeeper, and it handles the whole flow: shows products, negotiates the price, collects name, size, quantity and address, saves the order to a sheet, and looks up order status later. What it can do - Replies in Roman Urdu, answers questions about products, shop location and orders - Shows products 3 messages at a time (to save cost), then the next batch if they ask for more - Understands photos/screenshots of a coat the customer found elsewhere and checks if we have the same or a similar one - Handles voice notes (transcribed, then answered) Stack: n8n webhook, code node to drop old/duplicate events, photo/voice to text, AI agent with 3 tools (catalog search, save order, order lookup), a code check on the reply, then the reply is split into text and real images and sent. The problem It works perfectly when I test it, because I know how to order properly. With real customers it breaks. People type casually, don’t know waistcoat terms, and the agent misreads what they want. It sometimes sends wrong products, and after a customer picks one item, it still sends photos of the other products it showed earlier. What I’m unsure about - Is this a system prompt problem, a catalog search problem, or an architecture problem? - Should I add a step that rewrites the customer’s message into a clean query before the agent sees it? - How do you stop the agent from “forgetting” which product the customer picked? I can share the system prompt and a failed chat if that helps. The template is "Free" to grab. Go check it out and let me know what could be changed or improved...