AI WhatsApp Support Assistant for Al M***** Retail, UAECase Study

The large majority of incoming messages were variations of 'where is my order', which meant experienced agents spent their day copying tracking numbers instead of resolving genuine problems. The language requirement was the hard part: customers wrote in English, in Modern Standard Arabic, in Gulf dialect, and very often in a mix of the two within a single message, sometimes using Arabic written in Latin characters. Off-the-shelf intent models handled English acceptably and Gulf dialect badly. There was also a trust constraint — the client was clear that a bot which trapped customers in a loop would damage the brand more than slow replies did, so escalation had to be fast and obvious. On the technical side, the order data lived in the ecommerce platform with no clean API for status lookups, and WhatsApp's Business API imposes strict rules on what can be sent outside a 24-hour customer service window, which shaped what the assistant could proactively do.

We built the assistant as a Node.js service handling WhatsApp Business API webhooks, with a separate Python service doing intent classification and entity extraction. Rather than relying on a single multilingual model, we trained intent classification on the client's own historical WhatsApp transcripts, which is what made Gulf dialect and mixed-script messages work — the training data reflected how their actual customers write, not how a generic corpus does. Order lookups run against a small integration layer we built over the ecommerce platform, caching status in MongoDB so a lookup does not hit the store on every message. The assistant handles order status, delivery timelines, returns initiation and basic product availability; anything else, or any message where intent confidence falls below threshold, hands off to a human agent with the full conversation context attached. We deliberately made the handoff a single step rather than a menu tree. Arabic responses were written by a native speaker rather than machine translated, because a bot that replies in stilted Arabic reads as a foreign company. Message templates were pre-approved through WhatsApp's process to allow proactive shipping notifications within policy. We also built a small agent console in React so supervisors can watch live conversations, review what the assistant handled overnight, and correct misclassified intents directly — those corrections feed the next retraining cycle rather than sitting in a ticket queue. Related: AI & ML development and our ecommerce practice.