Grounding beats fluency

The assistant answers through OpenAI function-calling over real product data — search the catalog, read stock and price, return what actually exists. Fluency without grounding is a liability; a confident wrong answer costs a sale and trust.

Every customer-facing claim traces back to a catalog lookup, not to the model's imagination.

One assistant, multiple channels

Telegram and website chat share the same core, so the shop maintains one assistant, not one per channel. Vision and voice input are supported, because customers describe a watch with a photo or a voice note more naturally than with a SKU.

The channel is a thin adapter; the reasoning and the catalog access are shared.

A staged-rollout safety architecture

A new model or prompt does not go straight to customers. It moves through controlled stages — shadow, then a limited release — so behaviour can be observed before it is trusted at scale.

The rollout gates are the reason a shop owner is willing to leave a generative system switched on.

What it deliberately is not

It is not a general chatbot bolted onto a store, and it is not an autonomous agent making commercial decisions. It answers catalog questions and hands off when it should.

Scope discipline is what keeps a helpful assistant from becoming an unpredictable one.

Why this is the hard part

Most AI product tooling optimises for throughput. On a live storefront the constraint is trust: an assistant a shop owner is willing to keep running is worth more than one that talks more and grounds less.

The value is a catalog-aware sales channel released safely, not a demo.

For implementation detail, see AI Sales Assistant — showcase and AI sales consultant. You can also browse the project library and the case studies.