top of page
How a governed Fabric warehouse let a operations team 'ask their data' with an AI data agent — no SQL
Problem Statement
An operations team wanted business users to get answers ("Which regions missed SLA last quarter and
why?") without waiting in the BI backlog for an analyst to write SQL. They'd seen AI chat demos but
worried about two things: wrong answers on messy data, and AI touching data it shouldn't.
Why the warehouse comes first
AI data agents are only as trustworthy as the data underneath them. Pointing an LLM at raw,
ungoverned, duplicated data produces confident-but-wrong answers. A well-modeled Fabric warehouse
solves exactly this: it stores cleaned, conformed, business-ready data with a single source of truth,
consistent KPI definitions, and permissions — the substrate a data agent needs.
What we did (warehouse + data agent approach)
1. Built the governed Fabric warehouse first. Consolidated ERP/CRM/ops sources into OneLake,
modeled a star schema with clear, named measures (SLA %, on-time delivery, cost-to-serve) so an
agent has unambiguous definitions to reason over.
2. Enabled AI access over that governed layer. With everything in OneLake in open Delta format,
AI agents can explore schemas and tables directly. Microsoft's OneLake MCP (generally available,
2026) lets an AI agent discover workspaces, read table schemas, and answer questions about the data
through natural-language conversation — and a Fabric Data Agent can be published so business
users ask questions in plain English and get governed answers.
3. Kept it secure by design. OneLake AI tooling operates under the user's existing Azure identity
and Fabric permissions — "your AI agent can only access what you can access." No new data exposure;
row/column security and workspace roles still apply.
4. Grounded answers to curated measures. By restricting the agent to the modeled semantic layer
(not raw tables), answers stay consistent with the KPIs leadership already trusts.
Business Outcome
- Self-service answers in seconds for non-technical users - "one conversation" instead of a ticket and a multi-day analyst turnaround.
- Trustworthy, governed responses because the agent reasons over cleaned, conformed warehouse data with defined measures, not raw silos.
- No new security surface - the agent inherits each user's existing Fabric permissions.
- A reusable AI foundation - the same governed warehouse now powers dashboards, agents, and future ML, instead of a one-off chatbot.
Find Us Here
bottom of page