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.

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.