The problem
Support chatbots fail in two directions. One invents a policy that sounds right and costs a refund. The other is so hedged that every conversation ends in 'contact us'. Both of them end with a person reading the whole thread from the top.
What we built
The assistant answers only from the shop's help centre and names the article it used. Asked something the help centre does not cover, it says so. It looks up an order, but only the order belonging to the customer it is talking to, and checks return eligibility against the policy as written in code rather than against the model's memory of it. It can open a ticket. Before it writes a note to the customer record it asks, because writing to someone's record is a different kind of act from reading it. Contact and payment details are stripped out of a message before anything stores it, and it answers in the language the customer wrote in.
The result
Hearthline Kitchen is a fictional shop and the page says so. A person can take the thread over and reply from the inbox while the assistant stays quiet, and the owner can edit a help-centre article and watch the next answer change. That last part is what decides whether a support assistant stays maintainable by the team that owns the content.