Chow Tai Fook: 400+ AI agents for jewelry retail staff

Gold and jewelry products are crafted and far from standardized, customers need a high level of trust, and service depends on years of frontline experience. Store staff had to move between separate systems for inventory, pricing, gold prices and compliance, and the size of the store network made new ideas hard to roll out everywhere.¹

The company set a governed base first. Microsoft 365 E5 is its standard platform for collaboration and security, Microsoft Purview handles data governance and monitors unsanctioned AI use, and a unified security platform correlates signals across endpoints, identities, email and cloud apps.

On top sits "AI Fook", an ecosystem of agents built on Microsoft Foundry that lets staff ask in plain language about product stories, inventory and styling. Other tools include AI role-play training that turns top sales behaviors into standard procedures, an AI Insights Platform for managers, two-track approvals in which AI clears routine transactions and flags complex ones for people, vision models that watch high-risk store operations, an omnichannel engine that links online search with store visits, and Azure OpenAI Service (now Azure OpenAI in Microsoft Foundry Models) for 3D design concepts and content. Microsoft Fabric and GitHub Copilot are named without detail. Agent-native storefronts, agent-to-agent commerce and custom designs generated from a customer's story are described as exploratory.¹²

In our view, the order is the lesson: Chow Tai Fook put identity, security and data governance in place, including visibility of unsanctioned AI use, before rolling out hundreds of agents. Retailers planning many agents can reuse that sequence. The figures need caution. Both sources are Microsoft pages, no figure is attributed to a named person, the efficiency gain has no defined baseline, and the conversion number is a ceiling that should not be quoted as typical. Several features are still exploratory. A retailer with standardized products and one main system may need a few well-designed agents rather than hundreds.