A conversation that knows the catalog

The question search cannot answer

A shopper asking for a gift within a budget is not asking for a keyword match. They have constraints: who it is for, what they can spend and what is available. A useful retail conversation must stay attached to the systems that know those facts.

Built into the storefront

We built a commerce agent deployed on a large Southeast Asian retailer's site. It is part of the storefront rather than a separate chat surface. A shopper can ask for a budget-constrained gift, an available substitute or a compatible pairing. The agent presents product cards, comparisons or a basket panel inside the conversation.

Language is part of the product

The system supports Thai-first support alongside English and regional languages. Product copy is curated per language rather than treated as a machine-translation pass. That is a design choice about the whole buying interaction, not a claim about a measured translation benchmark.

Catalog, support and a handoff

The system is connected to the live catalog, product prices and stock, customer support, and point-of-sale information. Support questions can be answered in the same place or passed to a person with a conversation summary. When the agent is unsure, its behavior is to say so and hand over, not guess.

Three requests a search box cannot answer

The agent is built around the kinds of requests shoppers actually make. Three of them show why the catalog connection matters.

In each case the answer is only as good as the data behind it. A conversation that is not attached to the live catalog can sound helpful and still be wrong.

When the agent is unsure

The behavior we care about most is what the agent does when it does not know. It says so, and it hands the conversation to a person with a summary of what was said. The shopper does not have to repeat themselves, and the person picking it up does not start cold.

This matters more than fluency. An agent that guesses confidently is a liability in a store. An agent that knows its limits is a colleague.

What the storefront gains from this

Because the agent sits inside the storefront and not in a separate chat window, a shopper can move between a product card, a comparison and a basket without leaving the page. The conversation is part of the shopping, not a detour from it.

Deployment is not an outcome metric

The agent handles real production conversations. We are holding the client name and hard numbers until cleared to share. We can describe the deployed interaction and its system connections. We cannot claim a conversion increase, revenue gain, language accuracy rate or number of customers without a verified result.