The Architecture Behind Secure AI Analytics on a Data Lakehouse

Once your company data sits in a well-modeled lakehouse, AI assistants like Claude, ChatGPT, and Copilot can build board decks in your branding, run what-if scenarios on pricing and demand, and pull in outside context like weather to explain a dip in sales. In this episode, Brick Thompson and Landon Ochs walk through the architecture we use to connect those assistants to client data through a custom MCP server. Landon covers each step of the flow, starting with sign-in through Entra, where security groups determine which tools and data each person can reach. Business users get curated views with the company’s metric definitions built in, so a question like “show me sales” returns the right view for their role. Analysts and developers get read-only access across the bronze, silver, gold, and platinum layers to accelerate their Power BI and Excel work. Every query runs under the user’s own identity, which carries row level security through to the AI and records the person’s name in the audit log. The conversation also covers searching unstructured sources like Gong transcripts, PDFs, and emails through Azure AI Search, why midsize companies with multiple ERPs or recent acquisitions benefit from consolidating data before connecting AI, and how a custom MCP server can be repointed from Fabric to Snowflake, BigQuery, or Redshift when your platform changes. Blue Margin builds and manages data platforms for midsize and PE-backed companies, including the custom MCP servers that connect AI assistants to governed company data. Learn how we can help you put your data to work with AI: https://bluemargin.com