If you've ever watched a business analyst wait three days for a data team to answer a simple question like "What were our top 10 products by revenue last quarter?" — you know the pain. That lag between curiosity and insight is where decisions go to die.
Microsoft's Fabric Data Agent (currently in preview) is built to close that gap. It lets anyone in your organization ask questions about enterprise data in plain English and get structured, accurate answers — no SQL, no DAX, no KQL required.
Let's dig into what it actually is, how it works under the hood, and what you need to know to get started.
What Is a Fabric Data Agent?
A Fabric Data Agent is a conversational Q&A system built on top of your organization's data in Microsoft Fabric's OneLake. It uses large language models (LLMs) — specifically Azure OpenAI's Assistant APIs — to interpret natural language questions and translate them into queries against your actual data sources.
Think of it as a smart intermediary that:
Understands your question
Figures out which data source best answers it
Generates and executes the right query
Returns a human-readable answer
The agent can connect to lakehouses, warehouses, Power BI semantic models, KQL databases, ontologies, and Microsoft Graph — all within the governed Fabric ecosystem.
Within broader agentic architectures on Microsoft Fabric, data agents serve as the conversational analytics component, connecting to governed data through multiple data source types in multi-agent solutions.
How It Works: Under the Hood
The magic happens in a well-orchestrated pipeline. Here's the flow:
1. Question Parsing & Validation
When a user submits a question, the agent applies Azure OpenAI Assistant APIs to process it. Before anything else, it checks that the question complies with:
Security protocols
Responsible AI (RAI) policies
The requesting user's permissions
The agent operates with read-only access to all data sources — it cannot write, modify, or delete data.
2. Data Source Identification
The agent uses your credentials to access the schema of available data sources (not the data itself). It evaluates your question against all connected sources and determines which one is best positioned to answer it. You can even add custom instructions to guide this routing — for example: "Direct financial metric questions to the Power BI semantic model; route raw data exploration to the lakehouse."
3. Query Generation
Once the right data source is identified, the agent generates the appropriate query using one of these translation tools:
Data SourceQuery TypeLakehouse / WarehouseNL2SQL (Natural Language → SQL)Power BI Semantic ModelsNL2DAX (Natural Language → DAX)KQL DatabasesNL2KQL (Natural Language → KQL)Microsoft GraphGraph API queries
4. Query Validation & Execution
The generated query is validated for correctness and security compliance, then executed against the data source. Results are formatted into a human-readable response — tables, summaries, key insights — and returned to the user.
Configuring a Fabric Data Agent
Setting up an agent is described as being similar to building a Power BI report: you design and refine it, then publish and share it. Here's what configuration involves:
Select Your Data Sources
An agent supports up to five data sources in any combination — lakehouses, warehouses, KQL databases, Power BI semantic models, ontologies, or Microsoft Graph. You could have five Power BI semantic models, or a mix of two semantic models, a lakehouse, and a KQL database.
Choose Relevant Tables
After adding a data source, you define which specific tables the agent can access. For lakehouses, this means lakehouse tables (not raw files). If your data lives in CSV or JSON files, you'll need to ingest it into tables first to make it available to the agent.
Add Context with Instructions & Example Queries
This is where you fine-tune the agent for your organization:
Data agent instructions — Tell the agent how to behave. Define which data source to use for which type of question. Clarify organizational terminology. Set custom rules.
Example query pairs — Provide sample question-to-query mappings so the agent learns how to handle common queries in your domain. (Note: example query pairs aren't yet supported for Power BI semantic model sources.)
Security & Governance: Built-In, Not Bolted On
One of the more impressive aspects of Fabric Data Agent is how deeply governance is integrated:
Least-privilege access: The agent uses the requesting user's credentials, so it can only surface data that person is already authorized to see.
Microsoft Purview integration: DLP policies, access restriction policies, Insider Risk Management, and audit/eDiscovery all apply to agent interactions.
Guardrails on scope: Queries are constrained to configured data sources — the agent can't go rogue and query things outside its defined scope.
Optional Azure AI Content Safety: You can add an extra layer of content risk controls to filter harmful or out-of-policy responses.
Beyond the Chat Window: Copilot Studio Integration
Fabric Data Agents aren't limited to the Fabric portal. You can consume a Fabric data agent in Copilot Studio, embedding your data agent into custom Microsoft 365 Copilot experiences, Teams bots, or other applications. This opens the door to deploying data conversations wherever your users already work.
Why This Matters
The Fabric Data Agent addresses a real organizational problem: data insight accessibility. Most enterprise data is technically available but practically inaccessible to the majority of people who need it, because it requires technical skills to query.
By enabling plain-English conversations with governed data, the agent:
Lowers the barrier for non-technical stakeholders
Reduces the bottleneck on data teams for ad-hoc queries
Fosters a culture of data-driven decision-making
Keeps everything within your existing governance and security boundaries
This isn't a chatbot on top of a CSV. It's a governed, multi-source, enterprise-grade conversational analytics layer built into the same platform where your data already lives.
Top comments (0)