Overview
Most organizations want to put AI assistants like Claude, ChatGPT, and Copilot to work on their data, but few have a clear answer for how to do it without opening the door to security gaps, inconsistent answers, or ungoverned access. In this episode of The Dashboard Effect, Brick Thompson and Landon Ochs walk through the architecture behind secure AI analytics on a data lakehouse, from identity and access controls to the custom MCP server that connects AI tools to governed data.
The conversation is practical and specific, covering how each layer of the stack works together so that AI respects the same permissions, definitions, and audit requirements as the rest of the business. See how Blue Margin’s Managed Data Platform helps organizations give AI secure, governed access to the data that drives their decisions.
What This Episode Covers
Identity and Access with Entra (5:12 – 8:47)
The architecture starts with Entra authentication, which establishes who is asking a question before any data is touched. Security groups then define roles for business users, analysts, and developers, so each person reaches only the data that fits their responsibilities. This matters because AI assistants can surface information quickly and broadly, and without role-based boundaries in place, that speed becomes a liability. Brick and Landon explain why identity has to be the foundation rather than an afterthought.
Tiered Data Layers and User-Level Security (8:47 – 13:06)
The hosts break down how the lakehouse supports different levels of access, with curated Platinum layers for business users and raw Bronze and Silver layers for developers who need to work closer to the source. Every query runs under the user’s own identity, which means Row-Level Security carries over automatically and each request leaves a clear audit trail. For organizations worried about AI exposing sensitive records, this approach keeps existing governance intact instead of creating a parallel set of rules. It also gives leaders confidence that the answers people receive reflect the data they are authorized to see.
Bringing Unstructured Data into the Picture (13:06 – 14:01)
Structured data tells only part of the story, and this segment covers how Azure AI Search allows AI to draw context from sources like Gong transcripts, PDFs, and emails. Combining that content with lakehouse data lets AI connect what the numbers show with what customers and teams are actually saying. The result is analysis that reflects a fuller view of the business, which is often where the most useful insights come from.
ROI, Accuracy, and Context-Aware AI (14:01 – 15:17)
Brick and Landon make the case for routing AI through a governed lakehouse rather than connecting it directly to transactional systems. Consolidating data into a single source of truth improves accuracy and performance, and it avoids the inconsistent answers that come from AI pulling from scattered systems. The custom MCP server also embeds company-specific metric definitions and business logic, so AI understands internal terminology without constant retraining. This is where the business value of the architecture becomes clear: better answers with less ongoing effort.
A Platform-Agnostic MCP Server (15:17 – END)
The final topic focuses on the custom Model Context Protocol server that sits between AI assistants and the lakehouse. Because it is designed to be platform-agnostic, organizations can move between backends such as Fabric, Snowflake, BigQuery, or Redshift without rebuilding the entire integration. For companies making long-term platform decisions in a fast-changing market, that flexibility protects the investment and reduces the risk of being locked into a single vendor.
Who It’s For
This episode is worth your time if you are a data or IT leader evaluating how to introduce AI assistants safely, a CFO or operations executive who wants AI-driven insights without compromising governance, an analytics team member responsible for security and access controls, or anyone who is trying to move AI from experimentation to reliable, everyday use across the business.
Why It’s Worth a Listen
The most useful takeaway is that secure AI analytics is less about the AI tool and more about the architecture behind it. By running every query under the user’s identity and carrying Row-Level Security through to the AI layer, organizations can give people powerful new ways to explore data without rewriting their governance model. That principle alone addresses one of the biggest concerns leaders raise when AI comes up.
The discussion of the custom MCP server is also worth close attention. Embedding metric definitions and business logic into that layer solves a common frustration with AI, which is that it does not understand how a specific company defines revenue, margin, or customer churn. Pairing that with a platform-agnostic design means the work holds its value even as the underlying data platform changes.
For organizations still weighing whether to connect AI directly to source systems,
