AI support assistant for a growing e-commerce brand

Hiring seasonal agents meant training people for six weeks so they could help for eight, and quality suffered anyway: seasonal staff answered from incomplete knowledge, creating escalations that consumed the permanent team's time. In peak season, first response slipped past 24 hours and one-star reviews said so explicitly.

Roughly two-thirds of tickets were variations of the same three questions: where is my order, how do I return this, does this product fit my case, each answerable from data that already existed in Shopify and the brand's policy pages. The team's expertise was being spent on questions that didn't need it.

The founders had tested an off-the-shelf chatbot the previous year and pulled it within weeks: it answered confidently from nothing, invented a return policy, and generated more complaints than it resolved. Any new attempt had to be provably grounded before it faced a customer.

We built an assistant grounded exclusively in the brand's real policy pages, product catalog, and FAQ, with retrieval citations on every answer, so both the customer and the support team can see which policy an answer came from. Questions outside the approved knowledge get a polite handoff, never an improvisation.

The assistant connects to Shopify with scoped read permissions, so 'where is my order' gets a real tracking answer, and to the returns system with a narrow write permission that lets it initiate standard returns within policy. Anything nonstandard (damaged items, exceptions, disputes) routes to a human with the conversation summarized and the relevant order attached.

Escalation is confidence-based and deliberately conservative: below a threshold, the assistant hands off rather than guesses. We tested against a corpus of adversarial transcripts (angry customers, ambiguous requests, deliberate manipulation attempts) before the assistant met a real one.

Launch was staged behind the existing contact form for a month, answering in a review mode where agents approved outbound responses. Only after the approval rate held above 95% did it go fully live: first on chat, then on email.