Skip to content

Connect a Studio agent to a Microsoft Fabric data agent

Most agents answer from documents — this one answers from your data. Wire a Microsoft Fabric data agent into Copilot Studio so users can ask questions of your warehouse, lakehouse, and semantic models in plain language.

Stage: Copilot Studio · For: Maker, IT/Admin · Level: Advanced · Time: 40 min

Two products, two owners

A Fabric data agent is built and secured in Microsoft Fabric; you connect it from Copilot Studio. Steps and screens move on both sides — confirm against the Fabric data agent docs and the Copilot Studio knowledge sources docs.

Works on any harness — cost and coverage differ

This is a platform skill that applies whichever harness powers your agent. Keep one thing in mind: the standard and chat harnesses are covered by a Microsoft 365 Copilot license inside Microsoft 365 channels, while the GitHub Copilot harness bills Copilot Credits for all usage and a license never covers it. Compare the engines · estimate the net cost.

When to use this

Document grounding answers "what's our policy?" A Fabric data agent answers "what were last quarter's returns by region?" — questions whose answers live in structured data, not a PDF. When your users keep asking the analytics team for numbers that already sit in a Fabric warehouse, lakehouse, or semantic model, a data agent lets them self-serve in natural language, and Copilot Studio puts that in the flow of work in Teams.

This is a maker-and-IT job: the maker wires and shapes the conversation; IT/admin owns the Fabric data the agent is allowed to read.

What you'll need

  • A Microsoft Fabric data agent already built and tested in Fabric, over the data you want to expose.
  • Copilot Studio access in the right environment, with permission to add the data agent as knowledge.
  • Clarity on who can see what: the data agent answers with the user's permissions (or a fixed identity) — decide this deliberately before you publish.
  • Agreement with the data owner on scope — which tables/models are in, which are out.

Try it now — the prompt

Once the Fabric data agent is connected, shape how your agent uses it. Design the instruction so it routes data questions to Fabric and stays honest about limits:

When a user asks a question about [sales / inventory / returns] numbers, use the
connected Fabric data agent to answer from our [warehouse / semantic model].
Always state the time period and any filters you applied. If the data agent can't
answer or returns nothing, say so plainly — never estimate or invent a figure.
For questions outside the data (policy, how-to), use the other knowledge sources.

Why this works: it ties data questions to the Fabric data agent, forces the agent to surface the period and filters (so users can trust the number), defines the empty-result path, and keeps document-style questions on the document sources — four things that stop a data agent from confidently returning a wrong number.

Step by step

  1. Confirm the Fabric data agent works on its own. In Fabric, test it with the exact questions your users will ask. If it's wrong there, Copilot Studio won't fix it — the data agent is the engine.
  2. Add it as knowledge in Copilot Studio. In your agent, add the Fabric data agent as a knowledge source / connection and authenticate to Fabric.
  3. Decide the identity model with IT. Choose whether the agent queries as the signed-in user (results respect each person's data permissions) or a fixed identity. This is a governance decision, not a default — get it right before anyone outside the build team uses it. 
    Confirm whether this data agent answers using the end user's Fabric permissions
or a single service identity, and list exactly which datasets it can read.
  4. Write the routing instruction. Add the prompt above so data questions go to Fabric and everything else stays on your document sources.
  5. Test the hard cases. Ask a question it can answer, one just outside the data ("why did returns rise?" — analysis it can't do), and one a restricted user shouldn't see. Confirm honest answers and that permissions hold.

Screenshots

We deliberately don't ship screenshots that go stale — the Microsoft Copilot UI changes often. Follow the numbered steps above, which we keep current. Maintainers can regenerate fresh captures with the Playwright tool in tooling/screenshots/.

Make it better

A connected data agent is the start of a genuinely useful analytics assistant: - Blend data and documents. Pair the Fabric data agent with your policy docs so one agent answers both "what's the number?" and "what's the rule?" — and tells the user which source each answer came from. - Add an action on top. Once it can read a number, let it act — "if inventory for [SKU] is below threshold, draft a reorder request" via a connector or Power Automate flow. - Tighten the questions it handles. Use test cases to find where it guesses, then constrain the instruction so it routes those to a human instead.

📚 Learn more. The Fabric data agent documentation covers building and securing the data agent itself, and the Copilot Studio knowledge documentation covers connecting and grounding it.

Watch out for

  • The data agent owns correctness. Copilot Studio only routes the question; if the Fabric data agent returns a wrong number, your agent repeats it confidently. Validate the data agent in Fabric first, and keep validating it.
  • Permissions are the whole risk. A data agent can surface real business numbers. Decide the identity model deliberately and test with a restricted user — never assume the default is least-privilege.
  • It answers, it doesn't analyze. A data agent retrieves and aggregates; it's not a substitute for an analyst's judgement. Set that expectation so users don't treat "why" questions as gospel.

Where this leads

A data-grounded Studio agent is one of the strongest cases for the platform — self-serve answers over governed enterprise data. When the analytics get heavy enough that you need custom orchestration, evaluation gates, and your own code over Fabric, that's the signal to look at Stage 7 · Foundry. Until then, keep it in Studio where IT can govern it.

Was this page helpful?

Found an issue or want to suggest an improvement? Open a GitHub issue  ·  Star on GitHub