Fabric data agent: what it is, how to set one up and what to govern

A Fabric data agent is a configurable item in a Fabric workspace. You point it at data, add instructions, and colleagues then ask it questions in plain English. Microsoft's concept page says the agent decides which data source fits the question, writes a query, checks it, runs it and returns a readable answer, so the person asking does not need to write SQL, DAX or KQL.

Microsoft separates it from the copilots built into Fabric. The page says copilots come preconfigured, while a data agent is a standalone artifact you can tailor with your own instructions and example queries. It also says data agents can integrate with Microsoft 365 Copilot and can connect with external systems such as Microsoft Copilot Studio, Azure AI Foundry and Microsoft Teams.

Microsoft's how-to page describes the path. In your workspace, select New Item and search for Fabric data agent, then name it. The OneLake catalog opens so you can add data sources one at a time, up to five in any combination. In the Explorer you tick the tables the agent may use. For a lakehouse these are tables, not loose files, so CSV or JSON files must be loaded into tables first.

Next you add context. Data agent instructions are plain-English text, up to 15,000 characters on Microsoft's page, that say which source to use for which kind of question and define your business terms. Example queries are question and query pairs. Microsoft says these are not currently supported for Power BI semantic model sources. Then you test, and select Publish, which asks for a description. After publishing you have a draft version to keep refining and a published version for colleagues.

Microsoft also advises descriptive table and column names, because names such as SalesData and ActiveCustomer help the agent write better queries than TableA or C1.

Microsoft's governance section says the agent reads data under the requesting user's identity and permissions, and that configuring or sharing an agent does not grant access to its underlying sources. It also says Purview data loss prevention and access policies can limit what the agent returns, and that Purview features for agent prompts and responses, such as auditing, are described as being in preview.

That leaves choices that belong to your team. Which tables go in, since the agent can only answer from what you select. Who may publish, and who reviews instructions and example queries, since these shape every answer. Who checks sample answers against a trusted report before wider release. Microsoft itself suggests human escalation paths for sensitive requests and periodic review of instructions. Fabric also supports Git integration and deployment pipelines for agents, which gives you a way to test changes before production.