Guide · 7 min read
What is a Microsoft Fabric data agent?
"Can I just ask the data a question?" is now a fair request. A Fabric data agent makes it possible, provided the data underneath has been done properly.
Short answer
What is a Microsoft Fabric data agent?
A Fabric data agent is an AI assistant you set up over your own Fabric data, such as a lakehouse, warehouse or Power BI semantic model. People ask it questions in plain English, for example "which customers bought less this quarter than last?", and it writes and runs the query, then answers using your data.
Agents can be made available where people already work, including Microsoft Teams. They only see the data you point them at, and they respect the same permissions as the rest of Fabric.
How it works
What happens when someone asks a question
1. The question
A manager asks a question in plain English, in Fabric or in Microsoft Teams.
2. The agent plans
Using the instructions and example questions you have given it, the agent works out which tables and measures answer the question.
3. The query runs
It writes and runs the query against your governed data, within the asker's permissions.
4. The answer
It returns the answer, and the query it used can be inspected, so people can see where the number came from.
What it needs
The foundations that make answers trustworthy
- A Fabric capacity (F2 or larger) with Copilot enabled by your Fabric administrator
- Data modelled once, with clear table and column names
- Measures defined the same way as in your existing reports
- Instructions that explain your business terms, such as what counts as a sale or a margin
- Example questions with known correct answers to test against
From our delivery work
How we roll one out
For a furniture and fabrics manufacturer, we set up a data agent in Teams over their validated Fabric model so the finance and leadership team could ask questions directly. Before anyone relies on it, we test it against figures they already know from Power BI, add currency labels and business terms to its instructions, and tighten the prompts where answers are ambiguous.
The agent sits on top of the lakehouse and reports; it does not replace them. Dashboards answer the questions people ask every week. The agent answers the ones nobody built a report for.
Be careful of
Where data agents go wrong
Unvalidated data
An agent over raw exports will give confident answers that do not match finance's numbers. Validate the model first.
Ambiguous terms
If "sales" means invoiced revenue to finance and orders to the sales team, the agent needs to be told which.
No test set
Without known questions and answers to check against, nobody can tell whether the agent is right.
Questions we get asked
Frequently asked questions.
Is a Fabric data agent the same as Copilot?
They are related. Copilot in Fabric helps people build and explore. A data agent is set up over specific data with your own instructions and examples, so it answers business questions consistently.
Can we use a Fabric data agent in Microsoft Teams?
Yes. Agents can be made available to people in Microsoft Teams, so they can ask questions without opening Fabric or Power BI.
Does a data agent see all our data?
No. It only uses the data sources you add to it, and it respects the permissions people already have.
What Fabric capacity do data agents need?
Microsoft requires at least an F2 (or P1) capacity, with Copilot enabled by your Fabric administrator. Agent usage draws from your capacity, so heavy use may need a larger size.
Next steps
Want to ask your own data a question?
In a 30-minute call we'll look at whether your data is ready for a data agent, and what it would take to get there.
Related reading
- Microsoft Fabric consultants
- Copilot Studio AI agents
- What is AI automation?
- How to choose a Microsoft Fabric capacity size
Prefer to talk it through? A 30-minute call is usually enough.
Bring us the process that is slowing you down.
A 30-minute discovery call is enough to work out whether automation is the right answer, and what it would take.