[{"data":1,"prerenderedAt":103},["ShallowReactive",2],{"work-ecommerce-support-assistant":3,"work-next-\u002Fwork\u002Fecommerce-support-assistant":100},{"id":4,"title":5,"accent":6,"body":7,"client":61,"cover":62,"coverAlt":62,"description":63,"duration":64,"extension":65,"featured":66,"industry":67,"meta":68,"navigation":66,"order":69,"path":70,"quote":71,"results":75,"seo":85,"services":86,"stack":90,"stem":97,"year":98,"__hash__":99},"work\u002Fwork\u002Fecommerce-support-assistant.md","ChatGPT-powered support assistant for an online retailer","#0E7490",{"type":8,"value":9,"toc":53},"minimark",[10,15,19,23,39,43,46,50],[11,12,14],"h2",{"id":13},"challenge","Challenge",[16,17,18],"p",{},"A fast-growing retailer was answering thousands of order-status, returns and product questions a month with a small support team. Response times stretched past a day at peak, satisfaction was slipping, and the existing rules-based bot answered only exact keyword matches.",[11,20,22],{"id":21},"what-we-built","What we built",[24,25,26,30,33,36],"ul",{},[27,28,29],"li",{},"A retrieval layer over the help centre, policy documents and product catalogue, refreshed nightly.",[27,31,32],{},"A GPT-4o assistant with tool access to the order-management API, so it can look up shipments and start a return, not just describe how to.",[27,34,35],{},"Confidence thresholds and intent-based escalation into Zendesk with a full transcript and suggested reply for the human agent.",[27,37,38],{},"An evaluation set of 400 real questions scored every release, plus a dashboard for resolution rate, hand-off rate and cost per conversation.",[11,40,42],{"id":41},"results","Results",[16,44,45],{},"Within the first quarter after launch the assistant resolved roughly four in five conversations end-to-end, cut first-response times by more than half, and lifted CSAT on bot-handled conversations by around 40% versus the previous bot. Support headcount stayed flat through a peak season.",[11,47,49],{"id":48},"what-made-the-difference","What made the difference",[16,51,52],{},"Grounding in live order data (not just documents), and treating the human hand-off as a first-class feature rather than a failure state.",{"title":54,"searchDepth":55,"depth":55,"links":56},"",2,[57,58,59,60],{"id":13,"depth":55,"text":14},{"id":21,"depth":55,"text":22},{"id":41,"depth":55,"text":42},{"id":48,"depth":55,"text":49},"Online retailer · consumer electronics (EU)",null,"A GPT-based assistant grounded in order data and help-centre content that now handles the majority of first-line customer inquiries.","10 weeks to launch","md",true,"Retail & E-commerce",{},1,"\u002Fwork\u002Fecommerce-support-assistant",{"text":72,"name":73,"role":74},"The assistant handles the bulk of first-line questions and hands over cleanly when it should. Our team finally has time for the complex cases.","Head of Customer Support","Client (attribution on request)",[76,79,82],{"value":77,"label":78},"80%","of inquiries resolved without an agent",{"value":80,"label":81},"60%","faster first response",{"value":83,"label":84},"+40%","customer satisfaction (CSAT)",{"title":5,"description":63},[87,88,89],"ai-chatbots","rag-systems","generative-ai-llm",[91,92,93,94,95,96],"OpenAI GPT-4o","LangChain","Pinecone","FastAPI","Zendesk API","n8n","work\u002Fecommerce-support-assistant","2024","hI2Lms3NqOrPd_UR1wDf5WMv8xxqWkULXLh3xh85vG0",{"path":101,"title":102},"\u002Fwork\u002Fstock-market-deep-learning","Deep-learning market analysis for a fintech",1792973679704]