{"id":4707,"date":"2025-04-24T08:16:06","date_gmt":"2025-04-24T15:16:06","guid":{"rendered":"https:\/\/devblogs.microsoft.com\/semantic-kernel\/?p=4707"},"modified":"2025-04-24T09:50:22","modified_gmt":"2025-04-24T16:50:22","slug":"customer-case-study-microsoft-store-assistant-bringing-multi-expert-intelligence-to-microsoft-store-chat-with-semantic-kernel-and-azure-ai","status":"publish","type":"post","link":"https:\/\/devblogs.microsoft.com\/agent-framework\/customer-case-study-microsoft-store-assistant-bringing-multi-expert-intelligence-to-microsoft-store-chat-with-semantic-kernel-and-azure-ai\/","title":{"rendered":"Customer Case Study: Microsoft\u202fStore\u202fAssistant \u2014 bringing multi expert intelligence to Microsoft\u202fStore chat with Semantic\u202fKernel and Azure\u202fAI"},"content":{"rendered":"

\"Picture1<\/a><\/p>\n

Introduction<\/h1>\n

In October\u00a02024<\/strong> Microsoft replaced a legacy rule\u2011based chat bot on Microsoft\u00a0Store with Microsoft\u00a0Store\u00a0Assistant<\/strong>, powered by Azure\u00a0Open\u00a0AI, Semantic\u00a0Kernel<\/strong>, and real\u2011time page context. The transformation changed a scripted, button-driven experience into a conversation that comprehends the entire public Microsoft portfolio, including Surface and Xbox products, Microsoft 365 subscriptions, Azure services, and the Dynamics and Power Platform portfolio, and knows when to involve a human Sales Associate. Six months later, the assistant manages several millions of conversations annually, maintains a four-star satisfaction rating, and generates revenue exceeding 140 percent of its forecast.<\/p>\n

The challenge before October\u00a02024<\/h1>\n

The previous solution relied on a rule\u2011based bot that steered shoppers through rigid decision trees and routed them to live agents via button clicks. Maintaining that framework was costly: each content update required manual topic management with answers, and every change spawned more flows to maintain. Human agents were inundated with off\u2011topic or incomplete chats because the bot could not reason over Microsoft\u2019s vast and constantly changing web presence, hundreds of thousands of pages spanning consumer and commercial offerings, hardware, software, and cloud services. Iteration of the legacy bot lagged behind product launches and new information, customers rated the bot poor on CSAT measures, and operational expenses of the bot and for human agents were unsustainable.<\/p>\n

From launch to current state<\/h1>\n

At launch, Microsoft Store Assistant utilized Azure Open AI gpt-4o with four \u2018experts\u2019 defined as Semantic Kernel skills, with streaming responses to the frontend. In subsequent months, the team upgraded model versions of gpt-4o through a simple configuration switch in Semantic Kernel and Azure Open AI, while automated simulations and evaluations with AI Foundry demonstrated functional and safety performance. New experts were added for tasks such as identifying appropriate phone numbers for customer requests, facilitating additional human transfers, and managing conversation closure. Real-time page context was integrated, fetching and parsing top-ranked Microsoft pages relevant to the discussion, which provided real-time detailed intelligence, along with stock and price data from Azure AI Search reasoning over the product catalog. Additional features included out-of-the-box solutions like Azure Open AI prompt caching, Azure Content Safety, and new features from Azure AI Foundry, all contributing to a highly performant, practical and maintainable assistant for customers.<\/p>\n

Multi\u2011expert orchestration with Semantic\u00a0Kernel<\/h1>\n

Store Assistant uses a multi-expert type of orchestration workflow in Semantic Kernel. There is a main \u2018Coordinator\u2019 that works like a planner and it decides which experts to invoke for each conversational turn. Each expert leverages a defined set of enrichment plug-ins giving it additional context to address the current conversation turn. This approach consistently provides the exact tool calls and context desired, with extremely low deviation. It also allows for the Coordinator to ask clarifying questions to improve hit rates of which expert and queries to invoke to improve the answer.<\/p>\n

Coordinator agent<\/h2>\n

Semantic Kernel\u2019s planner runs a comprehensive system prompt that:<\/p>\n