Momentum One Academy · glossary
Data Agent
Definition
A Fabric data agent is a configurable, read-only conversational artifact that lets people ask natural-language questions about selected, structured data in Microsoft Fabric. It evaluates the question, chooses an appropriate configured source, generates a SQL, DAX, or KQL query, validates it, and returns a human-readable answer. The agent can use lakehouses, warehouses, Power BI semantic models, KQL databases, ontologies, and other supported Fabric sources. 1 2
Where it sits in Fabric
The data agent is a consumption layer above governed Fabric data; it does not replace the data source or the business logic in a semantic model. When a Power BI semantic model is selected, the agent reads the model’s schema and metadata and uses natural-language-to-DAX generation to query it. Read permission on the model is sufficient for agent queries; Build permission is not required for this interaction. 1 2
It is related to, but distinct from, Copilot for Fabric. Copilot is a broader set of workload-specific assistive experiences, while a data agent is a standalone artifact that you configure for a defined set of data, instructions, and questions. 1 4
When to use it
Use a data agent when business users need governed, repeatable answers to questions such as sales by region, service volume by month, or inventory by category without writing queries. It is a good fit for structured-data exploration and conversational access to shared definitions. It is not a substitute for advanced analytics, causal analysis, machine learning, or document search; current guidance also limits it to read operations. 1 2
Practical example
A retail team publishes a curated sales semantic model with certified measures and clear table and column names. It configures a data agent with that model and an instruction to use the model for financial metrics. A manager asks, “What were net sales by region last quarter?” The agent generates DAX against the model and returns an answer that follows the user’s permissions and any applicable row- or column-level security. 1 3
Production caution
Do not treat a fluent answer as proof of correctness. Table selection, model metadata, instructions, examples, permissions, and changing source schemas all affect results. Test representative questions, inspect generated queries and intermediate steps, and publish a read-only version only after review. Sharing the agent does not grant access to its sources: each consumer still needs the required data permission, and RLS/CLS continue to apply. Administrators should also review Copilot/Azure OpenAI tenant settings and any cross-region processing or storage implications before rollout. 1 3 5
Module 8 connections
In Module 8, treat the semantic model as the governed definition of business meaning, Copilot for Fabric as an assistive experience, and the data agent as a controlled conversational entry point to that meaning. Governance supplies the permissions, policies, and review practices that keep the experience within approved data boundaries. 1 4 5
Build these governed patterns in Momentum One Academy’s Fabric Production Workshop.
References
Plan your workshop