How to Set Up a Microsoft Fabric Data Agent

A Microsoft Fabric Data Agent connects your business users directly to structured enterprise data. They ask a question in plain English; the agent generates SQL, DAX, or KQL behind the scenes, runs it against your connected data sources, and returns an answer. No dashboard-building required, no query language needed. Setup takes under an hour once the prerequisites are in place: F2 or higher Fabric capacity, six admin tenant settings enabled, and at least one structured data source in OneLake.

Prerequisites

Two things must be in place before creating an agent.

Fabric capacity: Your organization needs F2 capacity at minimum (or Power BI Premium P1 or higher with Fabric enabled). F1 and free trial capacities do not support Fabric Data Agents. Check the Admin Portal under Capacity Settings if you are unsure which capacity your workspace uses.

Admin access: A Fabric or Power BI admin must enable six tenant settings before the feature functions. These live in the Admin Portal, accessible through the gear icon in the top-right of Fabric. Plan for up to one hour of propagation time after enabling each setting.

Structured data: You need at least one data source already available in Fabric. Supported types are Lakehouse tables, Warehouses, Power BI semantic models, KQL databases, Ontologies, and Microsoft Graph. Raw CSV files stored in a Lakehouse file store are not supported. Load them into Lakehouse tables first.

Step 1: Enable the Required Tenant Settings

Open Admin Portal > Tenant Settings and turn on the following:

1. Users can use Copilot and other features powered by Azure OpenAI -- required for all AI capabilities.
2. Capacities can be designated as Fabric Copilot capacities -- required to assign your capacity for AI workloads.
3. Data sent to Azure OpenAI can be processed outside your capacity's geographic region -- required if your capacity is not in the EU data boundary or the US.
4. Data sent to Azure OpenAI can be stored outside your capacity's geographic region -- same geographic condition as above.
5. Conversation history stored outside your capacity's geographic region -- required for persistent conversational context. History is retained for 28 days and users can clear it at any time.
6. Allow XMLA endpoints and Analyze in Excel with on-premises datasets -- found under Integration Settings. Required only if connecting Power BI semantic models as data sources.

Settings propagate within one hour. Once active, return to your Fabric workspace to create the agent.

Step 2: Create a Data Agent in Your Workspace

1. Open your Fabric workspace.
2. Select + New Item.
3. In the All Items tab, search for Fabric data agent.
4. Select it, give the agent a name, and confirm.

The OneLake catalog opens automatically. This is where you connect data sources.

Step 3: Add Data Sources and Select Tables

Each agent supports up to five data sources in any combination. From the OneLake catalog, select a data source and click Add. To add more after the first, use + Data source in the left Explorer pane.

After connecting a source, the Explorer pane displays its available tables. Use the checkboxes to specify which tables the agent can query. Tighter table selection tends to produce more accurate responses.

Table naming matters more than most admins expect. Microsoft's documentation notes that descriptive names like SalesRevenueByRegion produce more reliable generated queries than generic names like Table1. If your existing tables have generic names, creating views with descriptive names is a practical workaround.

The query language the agent uses depends on the data source:

• Lakehouses and Warehouses generate SQL
• Power BI semantic models generate DAX
• KQL databases generate KQL

For Power BI semantic models, users only need Read permission on the semantic model. Workspace membership and Build permission are not required, which simplifies sharing with non-technical stakeholders.

Step 4: Configure Instructions and Example Queries

Two configuration options improve accuracy.

Agent instructions (up to 15,000 characters): Natural language guidance written by you. Use this to define routing logic across data sources (for example, "for financial metrics use the semantic model; for raw transaction data use the lakehouse"), clarify domain-specific terminology and abbreviations, and set response format expectations.

Example queries: Provide sample question and SQL or KQL pairs per data source. The agent retrieves the three most relevant examples when answering a new question. Up to 100 example pairs per data source are supported. This option is not available for Power BI semantic models or Ontologies.

Both settings are optional, but agents without configuration struggle with ambiguous questions. Writing five to ten representative question-query pairs for common request types substantially improves first-run accuracy with no additional infrastructure cost.

Step 5: Test, Publish, and Share

The built-in chat interface lets you test the agent before publishing. Ask questions representative of what your team would send. The agent shows its reasoning and the underlying generated query for each response, making it straightforward to identify where instructions or examples need adjustment.

When the agent is behaving as expected, click Publish. Write a clear description covering the agent's purpose, the data it covers, and the question types it handles well. This description is used by Copilot Studio and other orchestration layers to invoke the agent correctly.

Publishing creates two separate versions: a live published version and an editable draft. Share the published version with colleagues. Consumers need Read permission on the agent and Read permission on the underlying data sources. Row-Level Security and Column-Level Security configured on Power BI semantic models still apply, even when accessed through the agent.

For teams that need version control and promotion across environments, Fabric's Git integration can track agent instructions, example queries, and data source selections. Deployment pipelines support promotion from dev to test to production workspaces.

Connecting to Copilot Studio (Preview)

If your organization uses Microsoft Teams, you can surface the agent through a Copilot Studio bot. This integration is in preview as of April 2026 and requires a Microsoft 365 Copilot license for each person building or managing agents in Copilot Studio.

Setup: in Copilot Studio, create or open an agent, go to Agents > + Add, select Microsoft Fabric, and connect the published Fabric data agent. Enable generative AI orchestration under the agent's Settings panel. Publish the Copilot Studio agent to your Teams channel.

One current constraint: the combined deployment cannot be published to Microsoft 365 Copilot. Teams is supported; M365 Copilot is not, as of April 2026.

Key Limitations

Response size cap: Responses are capped at 25 rows and 25 columns. The agent summarizes or truncates beyond that. For large-dataset exploration, this is a real constraint. Starting a new chat session rather than continuing a long thread produces better results, since prior conversation history consumes context.

English only: Questions, agent instructions, and example queries must all be in English.

Structured data only: The agent cannot query raw files. PDFs, DOCX files, and CSV files not yet loaded into Lakehouse tables are outside scope.

Read-only: Only SELECT-equivalent operations are supported. The agent does not write, update, or delete data.

Single region: The agent and its data sources must be in the same geographic region. A lakehouse in North Europe fails if the agent's workspace capacity is in France Central.

Five data source limit: You can connect up to five sources per agent. Organizations with many data silos may need multiple purpose-specific agents.

If your team works across multiple file types and needs answers without setting up Fabric infrastructure, tools like VSLZ handle this from a direct file upload with no workspace configuration required.

Summary

A Fabric Data Agent reduces the gap between enterprise data and the business users who need answers from it. The setup path is predictable: enable six tenant settings, create the artifact in a workspace, connect up to five data sources, write instructions and example queries, then publish and share. The Copilot Studio integration extends that access into Teams without requiring SQL or DAX knowledge on the consuming end. The main constraints to plan around are the 25-row response cap, the English-only requirement, and the five data source limit per agent.