A WhatsApp chatbot that resolves most order-status questions by itself

Business context

A nine-person online specialty tea retailer sells loose-leaf tea and brewing gear through its own store, and support runs almost entirely on WhatsApp. Repeat customers ask where an order is, how long delivery takes, and how returns work. Three people answer by hand, and every reply is written as if it might earn another order.

We built a WhatsApp chatbot for ecommerce order questions so the routine ones no longer wait in a queue. The scope was narrow on purpose: answer order status, delivery and returns questions quickly, and pass everything else to the team with the full conversation attached. It follows the same approach as our other chatbot work, and a similar build is written up in the support chatbot we made for a scheduling product.

The problem

Every message landed in one shared inbox. Three people watched it during office hours, and nobody watched it on a Sunday. A customer asking at 8pm on Saturday often waited until Monday morning for an answer they could have found in seconds.

The questions repeated. About seven in ten were about order status or delivery. Answering one meant opening the store admin, finding the order, copying a tracking link, and typing a reply. That took two to three minutes, and the queue doubled during promotions.

Returns questions were harder to shortcut. The published policy covered most of them, but staff retyped the rule from memory because opening the policy page was slower than recalling it. When the queue grew, replies got shorter, and the friendly tone that keeps customers coming back slipped.

What we built

We connected the assistant to two sources: the store's live order data and the delivery and returns policy the retailer already publishes. Nothing else.

When a message arrives, the assistant decides whether it is about an order, a delivery estimate, or a return rule. For order questions it asks for the order number or the phone number used at checkout, finds the order in Shopify, and replies with the real status and tracking link. It never invents a delivery date. If tracking has not moved in three days, it says so and offers to pass the message to a person.

For delivery and returns it answers only from the published policy. When a question falls outside that text, it stops. It collects the order details and hands the conversation to the team in the same WhatsApp thread, along with a short summary of what the customer wants.

The assistant cannot issue a refund, change a delivery address, or promise a date. Those decisions stay with people. We chose not to build a general assistant that answers anything, because a wrong answer about an order costs more trust than a slow one.

Why we kept a person in the loop

Some messages look routine but carry a complaint inside them. A late order from a customer who has already emailed twice is not a status question, even if it is phrased like one. The assistant routes any message containing a complaint, a refund request, or a repeat contact to a person before it replies.

Rollout

• Week 1: Reading three weeks of chat history to list every question type and how often it appeared.
• Week 2: Building the order lookup and the policy answers, then testing both against the history.
• Week 3: Running the assistant in shadow mode, drafting replies that a person sent or discarded.
• Week 4: Turning on automatic replies for the three supported question types and leaving the rest to the team.

Before and after

Step	Before	After
First reply	Median 3.5 hours	Median 40 seconds
Weekend cover	None until Monday	Status and policy questions answered all weekend
Handover	Customer repeated themselves	Full conversation passed with a summary
Coverage	Three people, office hours	Three people, plus the assistant at all hours

Results

By the end of the pilot, 68% of conversations were resolved without a person. Median first reply fell from three and a half hours to forty seconds, and the team saved about nine hours a week that used to go into status lookups.

The numbers we watched most closely were the failures. In the first monthly review we found three wrong answers, all of them delivery estimates where tracking had stalled. After we added the rule that stalled tracking always triggers a handover, the next review found none. Wrong answers are rarer than the queue we started with, but they matter more, so we keep reviewing them every month.

Lessons learned

• The assistant got better when we gave it fewer topics, not more. Our first version tried to cover promotions and stock questions as well as orders. It was confident and often wrong on the extra topics, so we removed them and the supported answers became reliable.
• We assumed customers would use the website chat widget. They did not. Every real conversation happened on WhatsApp, so we deleted the widget and stopped maintaining it.
• Weekend is not a separate support channel, it is the same one with nobody watching. Measuring replies by day of week showed a gap that monthly averages had hidden.