Bilingual WhatsApp Support Agent for an E-Commerce Retailer

Failure mode 01 Bilingual conversations on one thread

Customers switched between Arabic and English, sometimes within a single conversation. The support team needed agents who could handle both languages comfortably and respond in the language the customer was using. This constrained hiring, increased training time, and made consistent response quality harder to maintain across shifts.

Failure mode 02 No continuity between conversations

Each customer conversation started without context. If a customer followed up on a previous inquiry, the agent had to reconstruct the history from scattered WhatsApp threads. There was no reliable mechanism to carry relevant context — a prior order, an ongoing issue, a previous resolution — from one conversation to the next.

Failure mode 03 Routine questions consumed agent capacity

A large share of inbound messages were predictable: product availability, shipping timelines, return policies, order status. Each one required an agent to look up information, compose a response, and send it. The volume of these routine inquiries created a ceiling on how many meaningful conversations the team could handle, because every routine exchange consumed the same time as a complex one.

Failure mode 04 No graceful escalation path

When a customer's issue went beyond what the team could resolve — a damaged shipment, a payment dispute, a request that needed human judgment — there was no structured mechanism to identify that need early and transfer the conversation cleanly. Escalation happened when an agent decided a conversation needed attention, which was reactive and inconsistent.

Decision 01 Bilingual understanding as a baseline requirement

The agent was built to detect the language of each incoming message and respond in the same language. Arabic and English conversations could proceed naturally, and the system would maintain the language match across a single thread even if the customer switched mid-conversation. This removed the language constraint from the retailer's support hiring and put it into the system layer.

“Respond in the customer's language, not the system's default.”

Decision 02 Per-customer memory across conversations

The system maintained a persistent memory per customer, storing relevant conversation history, resolved issues, and open items. When a returning customer messaged, the agent had access to prior context rather than treating each conversation as a fresh inquiry. This reduced friction for repeat customers and removed the need for them to re-explain issues they had already raised.

“Context follows the customer, not the thread.”

Decision 03 Structured handoffs instead of ad hoc escalation

When the agent determined it could not confidently resolve a conversation, it transferred the full thread to a human agent with structured context attached: what the customer asked, what the agent had attempted, and why the conversation needed human attention. The human agent received the conversation with the information needed to pick up where the AI left off, instead of reading through the entire history to reconstruct the situation.

“Hand off with context, not with a blank slate.”

Decision 04 Catalog grounding to keep answers accurate

Responses about products were grounded in the retailer's product catalog data rather than generated from the model's general knowledge. The system looked up product information from the actual catalog before composing an answer, which kept responses accurate and aligned with what the retailer was actually selling — current stock, prices, and product details.

“Answer from the catalog, not from memory.”

Decision 05 Human takeover as a first-class feature

Human takeover was built into the workflow as an explicit system capability, not an afterthought. The agent could recognize when a conversation had moved beyond its scope and transfer it to the human team. From the customer's perspective, the transition was clean — the conversation continued in the same WhatsApp thread, and the human agent stepped in with the context already assembled.

“The machine, not the customer, should ask for a human.”