DeoDap runs more than 350 chats a day across a 50,000-SKU wholesale catalog

DeoDap's customer is not a consumer buying one wall hook. It is a small reseller, a dropshipper or a shop owner buying twenty SKUs at a time, and their questions are operational. The four most common openings in the account are "Where is my order", "What are the best sellers?", "I want to edit my order" and "I want to return my order": three of four are about a transaction that already happened.

The intent mix confirms it. Order enquiry ran to over 4,000 chats, order tracking to over 2,000, returns and refunds to over 1,500 and grievance to over 1,000, over 9,000 chats, 54.7% of the total. No D2C brand documented here carries that ratio.

The recorded themes name the real problems, and they are trust problems: friction around COD and legitimacy, support bottlenecks concentrated in refunds, returns and order lookup, and courier and delivery timing uncertainty driving repeat tracking requests.

The objections are sharper still. COD unavailable blocks purchase outright for buyers who want it. Refunds feel slow and unclear, producing repeated "where is my refund" and "no response" threads. And order status problems (a draft status, an order not appearing, a zero-order screen) create distrust and abandonment. At more than 350 chats a day, none of that is answerable by a human team.

What resellers actually asked

The design assumption was inverted from a typical D2C deployment. Instead of a discovery agent that also answers order questions, DeoDap needed an order-state agent that also does discovery.

The Website Agent was trained on the catalog structure and price tiers, and (the operative part) on the order lifecycle: dispatch status, partial shipment handling, the return position on damaged goods, and the free-shipping threshold.

Automation carries the multi-step work. A refund question is never one exchange; it is a status check, then a timeline, then often an escalation. Handling that as a continuous thread rather than three disconnected tickets is what made 77.30% resolution possible on this mix.

Deflection is deliberately not maximised. 22.70% of chats reached a human, which on a mix containing over 1,000 grievance chats and over 1,500 return requests is a resolution policy rather than a failure. A frustrated reseller should reach a person faster than someone browsing wall hooks.

77.30% is the lowest resolution rate of the fifteen brands here, and it is the right number for this account.

Design for the post-purchase majority. over 9,000 chats were order enquiry, tracking, returns or grievance. Most conversational commerce deployments optimise pre-sale and treat support as overflow. At marketplace scale on a low-value basket, the post-purchase queue *is* the deployment.

Tune resolution per intent, not globally. A 77.30% overall rate with fast escalation on refunds beats a 92% rate that traps an angry reseller in a loop. Grievance was instrumented as its own intent (over 1,000 chats) rather than left to arrive as an unclassified escalation.

Keep discovery working underneath. Even on a service-heavy mix, 18.03% of chats ended in an order and Product Discovery still ran to over 5,500 chats. On a 50,000-SKU catalog, "what are the best sellers?" is a real and frequent question. The agent cannot only be a queue.