AI Sales Assistant for a Custom Sports Apparel E-commerce Platform (Claud API + MCP)

A US custom sports apparel company sells team uniforms through a quote-and-design funnel rather than a simple checkout: every order needs product and package selection, colors, a roster, a design request and designer approval before it can be produced. Inbound leads arrived over web chat, SMS and email around the clock, but reps could only respond during US working hours, and the same customer often restarted the conversation on a different channel. Leads went cold waiting for a first response, reps spent time re-asking for details the customer had already given elsewhere, and design requests were created inconsistently or in duplicate.

An AI sales assistant built on the Claude API, running as a Node.js service on AWS EC2 (Nginx, Supervisor) and integrated with the company's Laravel platform. Architecture: a Model Context Protocol (MCP) tool server exposes product discovery, catalog search, package contents, quoting, order placement, customer history, design requests and HubSpot CRM as tools. The assistant calls them under JWT auth with the same role-based access rules as the platform, so it can only see and do what a rep with that role could. Channels: a web chat widget on the storefront plus an authenticated customer chat, SMS, and email from the assistant's own mailbox. A unified cross-channel message log gives it memory of the customer across channels, and a reply gate decides per message whether to answer, stay silent or escalate. Behaviour: qualifies the lead, recommends the best-fit package, resolves colors, collects roster and design details, creates the design request against the real API, follows up on open requests with awareness of the designer's latest reply, and hands off to the assigned rep via Slack when ownership rules say so. In-stock products can be ordered end-to-end in the conversation. High-stakes steps (pricing floors, duplicate design requests, STOP handling) are deterministic code, not prompt text.

First response to every inbound lead on web chat, SMS and email around the clock, including outside US working hours, without a rep involved. Design requests are created by the assistant directly in the production system with the customer's actual package, colors and roster details; in-stock products are ordered end-to-end in the conversation (live in production). Customers who switch channels no longer repeat themselves; reps receive a structured Slack hand-off with full context instead of a cold lead. Duplicate design requests within a conversation are blocked by a deterministic guard; follow-ups check the designer's actual last reply before nudging a customer. Management receives a daily cross-channel report (one card per customer across chat, SMS, calls, email and design-request threads).