Fabric Data Agents: Building an Agent That Talks to Your Data in Teams

Since September 2026, Fabric data agents can be published to Teams and Microsoft 365 Copilot through Copilot Studio. Here's what they are, how to set one up, a retail example, and the settings to check before you go live.

Microsoft Fabric Oct 23, 2026 9 min read Atılay Berber

It’s 8:45 and a regional manager wants to know which five stores lost the most revenue yesterday. Today that answer usually lives on page three of a Power BI report, behind two filters and a drill-through. Or worse, in a Teams message to the BI team.

Fabric data agents exist to close that gap. The user asks in their own words, and the agent writes the right query against the right source and brings back the answer. The reason they’ve become one of the most talked-about Fabric topics of 2026 is a recent change: these agents can now go straight into Teams and Microsoft 365 Copilot.

A quick note: I covered the big picture of Fabric and the FabCon Europe 2026 announcements in my earlier post Microsoft Fabric in 2026: What Changed After FabCon?. This one is only about data agents.

What is a Fabric data agent, and how is it different from Copilot?

A Fabric data agent is a conversational Q&A layer over your own data. It takes the user’s question, decides which source to use, generates a query in that source’s language (SQL for lakehouses and warehouses, DAX for semantic models, KQL for KQL databases) and returns the result with an explanation.

Three things set it apart from general Copilot experiences:

You choose the sources. An agent can use up to five: lakehouses, warehouses, KQL databases, Power BI semantic models, ontologies and Microsoft Graph. You also pick which tables it can see in each one.
You write the business rules. Instructions like “revenue means net sales excluding VAT” or “weeks start on Monday”, plus example question-and-query pairs, teach the agent to speak your company’s language.
It only reads. The agent connects to its sources read-only and runs with the user’s permissions. Row-level security and Purview policies still apply.

What changed: the door to Teams and Microsoft 365 Copilot

For a long time, data agents mostly stayed inside Fabric. In September 2026 Microsoft announced that the Copilot Studio integration is generally available, and repeated it at FabCon Europe. What’s different now:

You no longer add a data agent to Copilot Studio as a “connected agent”. You add it as a tool called Fabric IQ Data MCP, and the Copilot Studio agent decides when it needs Fabric data.
One Copilot Studio agent can combine several data agents with SharePoint knowledge, connectors and workflows. A single answer can include a number from Fabric and a procedure from SharePoint.
You can publish the agent to Teams and Microsoft 365 Copilot, which is where business users will actually meet it.
The Fabric IQ MCP server is also GA, so the same governed answers are available to MCP clients like VS Code, GitHub Copilot or Claude Code.

One naming trap: Copilot Studio also has a separate Fabric IQ tool, still in prerelease, that connects to a semantic model directly. The one that calls your published data agent, with its instructions and example queries, is Fabric IQ Data MCP. They are not the same thing.

Before you start: prerequisites

Capacity: a paid F2 or higher Fabric capacity, or Power BI Premium P1+ with Fabric enabled. You can experiment on a trial, but plan around paid capacity.
Tenant settings: the Fabric data agent setting is off by default and an admin has to enable it. Copilot and Azure OpenAI settings must also be on, and semantic model sources need the XMLA endpoint setting. Changes can take up to an hour to apply.
Capacities outside the EU and US: if your capacity sits outside the EU Data Boundary and the US (Turkey, for example), you need to enable cross-geo processing and storage for AI. That isn’t really a technical switch; it’s a data residency and privacy decision. Talk to legal and security first.
Copilot Studio side: people who build and manage agents need a Microsoft 365 Copilot licence and a user licence. Copilot Studio agents consume Copilot Credits, including during testing and evaluation, so budget for it from day one.
Same tenant, same account: the data agent and the Copilot Studio agent must live in the same tenant, and you sign in to both with the same account.

Step-by-step setup

In Fabric:

Create a new Data agent in your workspace.
Add your sources (up to five) and select only the tables you need. A few well-chosen tables beat a long list every time.
Write the agent instructions: terminology, calculation rules, and which source to use for which kind of question.
Add example queries for each source. Give it the 10 to 20 questions real users will ask, with correct SQL or DAX.
Test, check the answers, then publish with a clear description. Only published agents show up in Copilot Studio.

In Copilot Studio:

Select the arrow next to “New agent” and choose Agent. (Clicking “New agent” directly skips this path.)
Under Build > Tools > +, search for “Fabric”, add Fabric IQ Data MCP, and pick your data agent from the OneLake catalog.
Write the tool description: which questions it answers, which business area it covers and what data sits behind it. The orchestrator relies on this text to decide which tool to call.
Choose the authentication mode (more on that below), save, and test in the Preview and Evaluate tabs.
Turn on the Teams + Microsoft 365 channel and publish.

A practical tip: if the Copilot Studio agent never calls your data agent, the problem is almost always the description or instructions. If it calls it but gets the answer wrong, the problem is in Fabric: the instructions, the example queries or the model underneath.

A retail example: the store performance assistant

Picture an agent for a supermarket chain with three sources:

Sales semantic model: stores, products, categories, daily revenue and margin measures
Competitor price lakehouse: shelf prices collected by web scrapers
External data warehouse: market data such as the national retail sales index

Part of the agent instructions:

- "Revenue" means net sales excluding VAT, after returns.
- Weeks start on Monday. "Last week" is the previous Monday to Sunday.
- Always use the Sales semantic model for revenue, margin and store comparisons.
- For competitor price questions, only use the latest price from the last 7 days.
- For market comparisons, use the retail index and mention that it is monthly.

An example question-and-query pair for the competitor price lakehouse:

-- Question: "Which competitor has the cheapest 5L sunflower oil?"
SELECT TOP 1 competitor, price, price_date
FROM competitor_price
WHERE product_name LIKE '%sunflower oil%' AND pack_size = '5 L'
  AND price_date >= DATEADD(day, -7, CAST(GETDATE() AS date))
ORDER BY price ASC, price_date DESC;

Once it’s in Teams, a regional manager can ask something like “Which category lost the most margin in the Izmir region last week, and how do our prices compare with competitors in that category?” in a single message, and the agent pulls from two sources to answer.

What to check before you go live

User or Maker? Copilot Studio can call the data agent in two ways. In User mode, queries run with the permissions of the person chatting, so RLS applies to each person. In Maker mode, everyone sees data through the permissions of whoever built the agent. If the builder can see every region, so can everyone who reaches the agent in Teams. Make User your default and keep Maker for data that’s genuinely meant for everyone, with the data owner’s sign-off.

Answers are small. The docs cap an answer at 25 rows and 25 columns (the newer preview runtime returns a summary plus a preview of up to 1,000 rows). Data agents are for answering questions, not for exporting data.

Data can leave the boundary. Responses returned to Copilot Studio may be processed outside Fabric’s compliance boundary and geographic region, under Copilot Studio’s terms. Put that decision in writing before you expose tables that hold personal data.

Region alignment. An agent can’t query a source in a workspace whose capacity is in a different region from its own. Check this at design time if you run more than one region.

The biggest risk is the model itself. The agent is only as good as your definitions. If two departments calculate margin differently, it will pick one and answer with complete confidence. Measure names, descriptions and one agreed definition per metric are the real work in this project.

Where to start

Collect the 20 questions users ask the BI team most often. That list becomes both your scope and your test set.
Start with one business area, say sales, and one semantic model.
Add descriptions to measures and tables, and fix cryptic column names.
Write instructions and example queries, then measure accuracy against your test set.
Pilot in Teams with a small group, in User mode, using real user accounts.
Add new areas like finance or inventory as separate, well-described agents. A few focused agents route better than one agent that knows everything.

Wrapping up

Bringing data agents into Teams changes where BI meets the user: a chat window instead of a report page. But the technology is the easy part; the setup really does take minutes. The hard part is the definitions, permissions and testing discipline that make the answers trustworthy.

My take: in 2026, a good semantic model isn’t just a good report anymore. It’s the company language your AI will speak. What’s the first question you think your users would ask in Teams?

Sources

STAY IN THE LOOP

New guides, sorted by topic in a single email.

Every Monday, the week's new posts arrive grouped by topic. No new posts, no email.