What if your sales team could ask "What was our pipeline coverage ratio last quarter by region?" and get an instant, accurate answer — without opening Power BI, without knowing DAX, without filing a request with the BI team?
That is what Fabric Data Agents do. They are conversational AI systems that sit on top of your Fabric data and answer natural language questions with structured, accurate results. No coding required to build one. No AI expertise needed to use one.
This guide covers what Data Agents are, how they work under the hood, and how to build your first one in under an hour.
What Are Fabric Data Agents?
Fabric Data Agents — formerly known as AI Skills — are standalone conversational artifacts in Microsoft Fabric. They use generative AI to translate natural language questions into queries against your data, execute those queries, and return structured answers.
The agent supports four query translation modes:
- NL2SQL — Natural language to T-SQL queries against lakehouses and warehouses
- NL2DAX — Natural language to DAX queries against semantic models
- NL2KQL — Natural language to KQL queries against eventhouses and KQL databases
- Microsoft Graph — Queries against organisational data in Microsoft Graph
A single Data Agent can connect to up to five data sources in any combination. You could have an agent that queries a lakehouse for raw transaction data, a semantic model for business metrics, and a KQL database for real-time telemetry — all through the same conversational interface.
How Data Agents Work
When a user asks a question, the Data Agent:
- Parses the question — Understands the intent and identifies which data source can answer it
- Generates a query — Translates the natural language into SQL, DAX, or KQL
- Executes the query — Runs it against the connected data source with the user's permissions
- Formats the response — Returns a structured answer, often with a table or summary
The agent respects your existing security model. It only returns data the querying user is authorised to see. There is no elevation of privileges, no bypassing row-level security, no shared service account.
Custom instructions let you shape how the agent behaves. You can tell it:
- Always include period-over-period comparisons
- Default to fiscal year rather than calendar year
- Use specific terminology from your business glossary
- Never return individual customer data, only aggregates
Example queries help the agent understand the kinds of questions it will receive. You provide sample questions and the expected query patterns, which improves accuracy dramatically.
Step-by-Step: Building Your First Data Agent
Step 1: Create a Data Agent
In your Fabric workspace, click New → Data Agent. Give it a name and description that reflects its purpose (e.g., "Sales Analytics Agent" or "Finance Q&A").
Step 2: Connect Data Sources
Add up to five data sources. Supported source types:
- Lakehouse — For raw and curated data in OneLake
- Warehouse — For SQL-based analytics
- Semantic model — For business metrics and DAX measures
- KQL database — For real-time and time-series data
- Ontology — For semantic relationships (preview)
- Microsoft Graph — For organisational data
For most Power BI teams, the starting point is a semantic model. This gives the agent access to your existing measures, hierarchies, and business logic without you having to re-explain anything.
Step 3: Add Custom Instructions
This is where you make the agent yours. Write instructions that tell the agent:
- Domain context: "You are a sales analytics assistant for a B2B SaaS company."
- Data conventions: "Revenue is always reported in USD. Fiscal year starts in July."
- Response format: "Always include the time period in your answer. Show percentages to one decimal place."
- Guardrails: "Never show individual deal values. Only show aggregates of 5 or more records."
Step 4: Add Example Queries
Provide 5-10 example questions with the expected query patterns:
- Question: "What was total revenue last quarter?"
- Expected pattern: DAX query against the Sales measure filtered by fiscal quarter
- Question: "Which region has the highest win rate?"
- Expected pattern: DAX query calculating win rate by region with a TOPN filter
These examples train the agent on your data vocabulary and query patterns.
Step 5: Test and Iterate
Use the built-in test chat to ask questions and verify answers. Common issues to watch for:
- Ambiguous terms: If "revenue" could mean gross or net, clarify in custom instructions
- Missing relationships: If the agent cannot join two tables, add the relationship to your semantic model
- Incorrect aggregations: If the agent sums when it should average, add example queries showing the correct pattern
Step 6: Publish
Once you are satisfied with the test results, publish the Data Agent. It becomes available in the Fabric portal, and you can configure access permissions.
Integration with M365 Copilot and Teams
The real power of Data Agents is in distribution. Once published, your agent can be surfaced in:
- Microsoft 365 Copilot — Users ask data questions in Copilot Chat, Word, Excel, PowerPoint, Teams, or Outlook
- Teams — The agent appears as a bot in Teams channels and chats
- Copilot Studio — Embed the agent in custom business applications
- Azure AI Foundry — Use the agent as a component in multi-agent orchestration
This means your sales team can ask pipeline questions directly in Teams. Your finance team can query budget data from Excel. Your executives can ask for revenue summaries in Outlook. No one needs to open Power BI.
Prerequisites for M365 integration: M365 Copilot license (separate from Fabric capacity) and Fabric data agent publishing permissions.
Governance with Microsoft Purview
Data Agents integrate with Microsoft Purview for enterprise governance:
- Data Loss Prevention (DLP) policies control what data can leave the organisation
- Access restrictions limit who can create and use Data Agents
- Risk discovery identifies agents that may expose sensitive data
- Auditing tracks all queries and responses for compliance
- eDiscovery includes Data Agent interactions in legal holds
- Insider risk management monitors for unusual query patterns
For regulated industries — financial services, healthcare, government — Purview integration is what makes Data Agents viable in production.
Real-World Use Cases
Sales Operations
Connect a Data Agent to your CRM semantic model. Sales reps ask: "What is my pipeline coverage for Q3?" or "Which deals are at risk of slipping this month?" The agent queries the semantic model and returns answers with the rep's own data filtered by their permissions.
Finance
Connect to your financial data warehouse. Finance teams ask: "What was our operating margin trend over the last 6 months?" or "Show me budget vs actual by department." The agent translates to SQL, executes against the warehouse, and returns formatted results.
Operations
Connect to your real-time telemetry in a KQL database. Operations teams ask: "Are there any anomalies in today's error rates?" or "What is the current throughput for the production line?" The agent translates to KQL and queries the eventhouse.
Customer Success
Connect to your customer health semantic model. CSMs ask: "Which accounts have declining usage?" or "What is our NPS trend by segment?" The agent provides instant answers without waiting for a scheduled report.
ALM Support: Git and Deployment Pipelines
Data Agents support Application Lifecycle Management:
- Git integration — Version control your agent configuration, custom instructions, and example queries
- Deployment pipelines — Promote agents from development to test to production
- CI/CD — Automate agent deployment as part of your Fabric workspace deployment process
This is critical for enterprise adoption. Your Data Agent is not a one-off experiment — it is a governed, versioned, deployable artifact that follows the same lifecycle as your reports and data pipelines.
Licensing
| Requirement | Details |
|---|---|
| Fabric capacity | F2+ or Power BI Premium P1+ |
| Data Agent creation | Pro or PPU license with workspace access |
| Data Agent usage | Any user with workspace access |
| M365 Copilot integration | M365 Copilot license (separate) |
| Purview governance | Microsoft Purview license |
Conclusion
Fabric Data Agents are the bridge between your Power BI data and the people who need it. They do not replace Power BI reports — they extend their reach to every user who has a question but does not have a dashboard open.
The barrier to entry is low: one hour to build, test, and publish your first agent. The impact is high: every user in your organisation gets a personal data analyst they can talk to in plain English.
Start with your most-requested semantic model. Build an agent. Give it to 5 users. Watch what happens.
Ready to build your first Data Agent? Book a discovery call with powerbi.ai and we will have your team talking to data within a week.