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

The large majority of incoming messages were variations of 'where is my order'. Experienced agents spent hours copying tracking numbers into replies instead of resolving genuine problems, and response times stretched during busy periods. Language made automation genuinely hard. Customers wrote in English, Modern Standard Arabic and Gulf dialect, frequently mixing them within a single message and sometimes writing Arabic in Latin characters. Generic intent models coped with English but handled Gulf dialect poorly.

There was a firm brand constraint as well. The client was clear that a bot which trapped customers in loops would do more harm than slow replies, so any escalation to a human had to be immediate and obvious. Arabic replies also needed to read naturally, because stilted machine-translated Arabic signals a foreign company to customers in the UAE.

The technical landscape added further limits. Order data lived inside the ecommerce platform with no clean API for status lookups, so the assistant had no direct source for its most common answer. WhatsApp's Business API also enforces strict rules on what businesses may send outside the 24-hour customer service window, which constrained any proactive messaging such as shipping updates.

PapaSiddhi built the assistant on the official WhatsApp Business API, with a Node.js service handling webhooks and conversation flow and a separate Python service performing intent classification and entity extraction. Instead of relying on one generic multilingual model, the team trained intent classification on the client's own historical WhatsApp transcripts. Because that data reflected how their customers actually write, Gulf dialect, Arabizi and mixed Arabic and English messages were handled reliably.

Order lookups run through a lightweight integration layer built over the ecommerce platform's API, with status cached in MongoDB so the store is not queried on every message. The assistant covers order status, delivery timelines, returns initiation and basic product availability. Anything outside that scope, or any message where confidence falls below a threshold, hands off to a human agent in a single step with the full conversation attached. A native speaker wrote the Arabic response library, and message templates were pre-approved through WhatsApp's process to allow proactive shipping notifications within policy.

Supervisors work from a React agent console where they can watch live conversations, review what the assistant handled overnight and correct misclassified intents directly. Those corrections feed the next retraining cycle, so the model keeps improving from real usage rather than leaving errors in a ticket queue. Supervisors can also see at a glance which conversations were escalated and why, helping the team spot new query types worth automating next.