Fabric Data Agents in Copilot Studio: Setup Guide

Navy Solv Systems title card reading 'Fabric Data Agents in Copilot Studio' with a laptop and cloud motif.

In Short: The Fabric IQ Data MCP Tool Is Now the Way In

Fabric data agents in Copilot Studio are generally available, and the way you connect them has changed. You no longer add a data agent as a "connected agent". You add a tool called Fabric IQ Data MCP to a Copilot Studio agent built on the GitHub Copilot harness, pick your published data agent from the OneLake catalog, write a description, choose an authentication mode, and publish to Teams and Microsoft 365 Copilot.

The data agent still runs in Fabric and still applies the permissions on its data sources. Copilot Studio decides when to call it. That split is the whole point: one governed definition of your numbers, reused by every agent that needs them. Most of the decisions that matter come down to two settings, the description and the authentication mode, and that is where I would spend your time.

Fabric Data Agents in Copilot Studio: What Changed at GA

The preview route treated a Fabric data agent as a connected agent: you went to Agents, chose Microsoft Fabric and selected it. That page is still on Learn, still labelled preview, and it states that a custom agent with a connected Fabric data agent is not supported in Microsoft 365 Copilot and is only validated for Teams.

The generally available route, announced on the Fabric Updates blog in September 2026 and repeated at FabCon Europe, is tool-based. The Copilot Studio agent calls the data agent the same way it calls any other tool, and the orchestrator decides when to reach for Fabric data and when to use its other knowledge and tools. You can add several data agents to one Copilot Studio agent, next to SharePoint knowledge, connectors and workflows.

It also widens the reach. Microsoft says you can now publish the Copilot Studio agent to Teams and to Microsoft 365 Copilot, which is where most business users will actually meet it. This builds on what we described when data agents were first exposed as MCP servers: the data agent stops being a destination and becomes a capability other agents consume.

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

Before You Start: Prerequisites

Microsoft's documentation lists these for the tool-based route:

Fabric capacity: a paid F2 or higher, or Power BI Premium P1 or higher with Fabric enabled
Tenant settings: cross-geo processing and storing for AI enabled where your region requires it
A working, published data agent: only published agents appear in the picker, and it should be published with a detailed description. Our guide to building your first Fabric data agent covers the build
Same tenant, same account: the data agent and the Copilot Studio agent must be in the same tenant, and you sign in to both with the same account
Permissions: at least read access to the data agent and its underlying sources, plus permission to create and modify agents in Copilot Studio
Licensing: Microsoft lists a Microsoft 365 Copilot licence and a user licence for each person who builds and manages agents

Budget for the Copilot Studio side too. Agents on the GitHub Copilot harness use Copilot Credits, and Microsoft says usage-based billing applies to using, building, testing and evaluating them. Our Copilot pricing and credits explainer covers how those credits are charged.

The Setup, Step by Step

This follows Microsoft's Learn article for the tool route:

Create the right kind of agent: in Copilot Studio, select the arrow next to New agent and choose Agent under "Optimize your business processes". This creates an agent on the GitHub Copilot harness. Selecting New agent directly skips the choice, and the steps do not apply
Add the tool: on the Build tab, go to Tools, select +, search for "Fabric" and choose Fabric IQ Data MCP. Create a connection if you do not have one
Pick the data agent: the OneLake catalog opens with the published data agents you can access
Write the description: say which questions it answers, the business area it covers and the data behind it. The orchestrator uses this to decide which tool to call
Choose the authentication mode: User or Maker, covered below
Add instructions: the agent's role, tone, and which data agent to use for which kind of question
Save: Microsoft is explicit that unsaved tools, descriptions and authentication settings are not applied when you test or publish
Test and evaluate: ask real questions on the Preview tab, then run a question set on the Evaluate tab before and after every change
Publish: select Publish, turn on Teams + Microsoft 365, choose "Make agent available in Microsoft 365 Copilot", then Save and publish

If the agent does not call your data agent during testing, the fix is almost always the description or the instructions, not the data agent itself. If it calls the data agent but the answers are wrong, the problem is in Fabric: instructions, example queries or the model underneath. We cover a wider testing approach in evaluating AI agents in Foundry and Copilot Studio.

User or Maker: The Decision That Matters

This setting decides whose access is used when Copilot Studio calls the data agent.

User: the credentials of the person chatting. Each user needs access to the data agent and its sources. People without access get no answer from it. Row-level and column-level security apply per person, as they do in Fabric
Maker: the credentials of the person who set up the Copilot Studio agent. Users need no access of their own, and everyone sees data through the maker's access

Maker mode is convenient, and that is the danger. If the maker can see every region's figures, so can everyone who can reach the agent in Teams. It quietly bypasses the row-level security you built in the model. Our default is User. We only use Maker for data that is genuinely meant for everyone who can reach the agent, with that decision written down and approved by the data owner. If you have not tested your security under conversational access, our walkthrough of the Fabric security model is the place to start.

Limits and Compliance to Check

Data residency. Microsoft warns that responses returned by the data agent might be sent outside the Fabric compliance boundary or geographic region, and are processed or stored under Copilot Studio's terms. For regulated UK and South African clients this needs a decision from whoever owns data protection, before go-live.

Answer size. Learn documents a cap of 25 rows and 25 columns per data agent answer. Microsoft's September 2026 feature summary says the preview runtime returns a summary plus up to 1,000 rows. A data agent answers questions. It is not a way to extract data.

Scope. One data agent can combine up to five data sources. If your Copilot Studio agent needs finance and sales, two well-described data agents usually route better than one agent stretched across both. That is the same logic as in our comparison of Fabric data agents and Copilot.

Region. A data agent cannot query a source whose workspace capacity is in a different region from the data agent's capacity. Check this before you design a cross-region setup.

Where Solv Systems Comes In

The connection itself takes minutes. Making it trustworthy takes longer: a data agent whose answers hold up, descriptions that route correctly, authentication that does not widen access, and an evaluation set that catches problems before your users do. Our AI automation and agent consultants build data agents in Fabric, publish them through Copilot Studio, and test them with real user accounts before anything reaches Teams.

Sources and Further Reading

Frequently asked

Yes. Microsoft announced general availability on the Fabric Updates blog in September 2026 and repeated it at FabCon Europe 2026. The generally available route is the Fabric IQ Data MCP tool in agents built on the GitHub Copilot harness. The older connected-agent route is still documented on Learn as a preview.

A paid F2 or higher Fabric capacity (or Power BI Premium P1 or higher with Fabric enabled), the cross-geo AI tenant settings where your region needs them, a published data agent with a detailed description, both products in the same tenant and signed in with the same account, and the Copilot Studio licensing Microsoft lists for the people who build and manage agents.

User authentication for almost everything. Each person then needs their own access to the data agent and its sources, and row-level security keeps working per person. Maker authentication runs every query with the agent author's access, so everyone using the agent sees what the author can see. Use it only where that is a deliberate, approved decision.

Yes, on the generally available route. You turn on the Teams and Microsoft 365 channel and select the option to make the agent available in Microsoft 365 Copilot. The older connected-agent documentation says Microsoft 365 Copilot is not supported, which is one more reason to build new agents on the tool-based route.

It can. Microsoft's documentation warns that responses returned by the data agent might be sent outside the Fabric compliance boundary or geographic region, and are then handled under Copilot Studio's terms and data handling policies. Check that against your data residency obligations before you publish.

Learn documents a cap of 25 rows and 25 columns per answer. Microsoft's September 2026 feature summary says agents switched to the preview runtime can return a summary plus a tabular preview of up to 1,000 rows. Data agents are built for answers, not bulk extracts.