Microsoft Fabric data agents: F2 capacity, licenses and cost per question
A Microsoft Fabric data agent needs a paid Fabric capacity of F2 or larger, or a Power BI Premium P1 or larger capacity with Fabric enabled, plus the Azure OpenAI tenant settings switched on by a Fabric administrator. There is no separate per-user data agent license inside Fabric. Every question is billed as capacity units (CU) against a Fabric capacity at 100 CU seconds per 1,000 input tokens, 10 CU seconds per 1,000 cached input tokens and 400 CU seconds per 1,000 output tokens, as listed on Microsoft Learn on November 12, 2026.
Microsoft describes the data agent as a generally available Fabric feature that lets teams build their own conversational question-and-answer system over data in OneLake. An analyst or BI developer picks up to five data sources, selects the tables the agent may use, writes instructions and example queries, and publishes the agent so colleagues can ask questions in plain English. The agent translates each question into SQL, DAX or KQL, runs it with the permissions of the person asking, and returns the answer.
Licensing changes when the agent is used outside Fabric. Publishing it to Microsoft 365 Copilot or connecting it to Copilot Studio is still in preview, and those routes add Microsoft Copilot or Microsoft 365 Copilot license requirements on top of the Fabric capacity.
··········
FABRIC DATA AGENT REQUIREMENTS: CAPACITY, SETTINGS AND DATA SOURCES.
Microsoft's "Fabric data agent concepts" and "Create a Fabric data agent" pages list the same prerequisites. The workspace that holds the agent must run on a paid F2 or higher Fabric capacity, or on a Power BI Premium per capacity SKU of P1 or higher with Microsoft Fabric enabled. A Power BI Pro or Premium Per User workspace on shared capacity does not meet the requirement, and the documentation does not list the Fabric trial among the supported capacities.
The agent also needs at least one data source that contains data and to which the creator has read access. The supported sources are a lakehouse, a warehouse, a Power BI semantic model, a KQL database (including Eventhouse), a mirrored database and an ontology. One agent can combine up to five sources in any mix, for example two semantic models, one lakehouse and one KQL database.
No Azure OpenAI key, service principal or API token is needed for the in-product chat, because Fabric uses a Microsoft-managed Azure OpenAI deployment that the customer cannot change.
........
Requirement | What Microsoft documents |
Capacity | Paid F2 or higher, or P1 or higher with Fabric enabled |
Data sources | Lakehouse, warehouse, semantic model, KQL database, mirrored database, ontology |
Sources per agent | Up to 5, any combination |
Agent instructions | Up to 15,000 characters |
Example queries | Up to 100 per data source; not for semantic models or ontologies |
Result size | Answers capped at 25 rows and 25 columns |
Languages | English only for now |
Model choice | Fixed by Microsoft |
........
The 25-row and 25-column cap shapes what the agent is good for. It suits questions such as total sales in a state for a given year, and Microsoft states plainly that it does not perform causal inference, so a question about why productivity fell in a quarter is out of scope. Unstructured files such as PDF, Word or text documents are not supported, and CSV or JSON files in a lakehouse have to be loaded or exposed as tables before the agent can read them.
··········
FABRIC DATA AGENT TENANT SETTINGS AND REGIONS.
The "Configure Fabric data agent tenant settings" page lists four settings in the Copilot and Azure OpenAI Service group of the admin portal. The first, "Users can use Copilot and other features powered by Azure OpenAI," is required for the data agent to work and can be managed at tenant and capacity level. The second, "Capacities can be designated as Fabric Copilot capacities," lets capacity admins create a dedicated capacity for Copilot and data agent usage.
The two remaining settings allow data sent to Azure OpenAI to be processed and stored outside the capacity's geographic region, compliance boundary or national cloud instance. Microsoft marks them as required when the capacity sits outside the EU Data Boundary and the United States. The storage setting also covers conversation history, which Fabric keeps for up to 28 days unless the user clears the chat. Microsoft notes that setting changes can take up to one hour to apply.
Regions also affect whether queries run at all. The documentation says a query fails when the data source's workspace capacity is in a different region from the data agent's workspace capacity, and gives the example of a lakehouse in North Europe and an agent on a capacity in France Central. Teams that spread workspaces across several regional capacities should place each agent next to its data.
··········
WHO NEEDS WHICH LICENSE FOR A FABRIC DATA AGENT.
Inside Fabric, access is governed by permissions on the data. The agent always runs queries and reads schemas with the credentials of the person asking, so each user sees only the tables and rows they are allowed to see. For a Power BI semantic model, Read permission on the model is enough to chat with an agent built on it; Build permission and workspace membership are not required, and row-level and column-level security still apply. Sharing an agent does not grant access to its underlying data, and Purview data loss prevention policies can block or trim answers.
The Fabric licensing page adds a rule that buyers should keep next to these permissions. A Fabric free license lets a user create and share non-Power BI Fabric items in a workspace on an F capacity, while viewing Power BI content on F SKUs smaller than F64 still requires Pro, Premium Per User or an individual trial. The data agent pages do not say whether chatting with an agent built on a semantic model counts as viewing Power BI content, so teams on F2 to F32 should confirm that point with Microsoft before rolling the agent out to free users. The comparison of Power BI Pro, Premium Per User and Fabric F64 sets out where that F64 line sits in the overall license plan.
........
Where the agent is used | Status | Licenses and access documented |
Fabric portal chat | GA | Read access to the agent's data; Read on semantic models |
Microsoft 365 Copilot (Agent Store) | Preview | Microsoft Copilot license or Office 365 commercial subscription, plus a user license per person |
Copilot Studio connected agent | Preview | Microsoft 365 Copilot license and a user license per builder |
All routes | n/a | Paid F2+ or P1+ capacity behind the agent |
........
The Microsoft 365 Copilot route requires the agent and Copilot to sit in the same tenant, with the user signed in to both with the same account. Publishing uses the "Publish to Agent Store" option, and recipients need access to both the agent and its data sources. Microsoft warns that the Copilot orchestrator reasons over the agent's output, so answers can be rephrased, and that responses sent to Copilot or Copilot Studio may leave Fabric's compliance boundary and fall under those services' own data handling terms.
In Copilot Studio, the builder chooses either user authentication, where each user needs access to the agent and its sources, or agent author authentication, where the author's credentials are used. That second option deserves a governance review, because every user of the custom agent then queries data through the author's identity.
··········
[[ADPLUS_300x250_1]]
HOW FABRIC DATA AGENTS CONSUME CAPACITY UNITS.
Microsoft's "Data agent consumption" page bills agent activity under the "AI Query" operation in the Capacity Metrics app, with usage grouped under the item kind LlmPlugin. Input tokens include the user's question plus the agent instructions, data source instructions, example queries and conversation history, which means long instructions and long chat sessions raise the cost of every question.
The SQL, DAX or KQL query the agent generates is billed separately to its own engine. A query against a warehouse, for example, appears under the SQL Query operation, so the true cost of a question is the AI Query charge plus the cost of the query it runs.
........
Token type | Rate per 1,000 tokens |
Input prompt | 100 CU seconds |
Cached input prompt | 10 CU seconds |
Output completion | 400 CU seconds |
........
Data agent work is classified as background activity, and Microsoft states that once the capacity is exhausted all operations shut down, so heavy agent use can block reports and pipelines on the same capacity. Customers pay the price of their billing region even when a request is processed elsewhere, and Microsoft can change the rates with notice by email or in-product notification.
For workspaces on capacities smaller than F64, the Fabric Copilot capacity page lists data agents among the supported experiences. When a capacity admin assigns users to a Fabric Copilot capacity, their data agent usage is charged to that capacity instead of the one holding the content. The Copilot capacity must be at least F2 or P1 and must sit in the tenant's home region, and each user can belong to only one. The Copilot in Power BI requirements cover the same mechanism for Copilot in reports.
··········
WORKED CAPACITY EXAMPLE FOR A FABRIC DATA AGENT ON F2, F4 AND F8.
Microsoft's own example uses 2,000 input tokens and 500 output tokens, which gives (2,000 × 100 + 500 × 400) / 1,000 = 400 CU seconds per request. The illustrative scenario below assumes a longer prompt, because instructions, example queries and chat history are counted as input: 3,000 input tokens and 600 output tokens per question, 25 analysts and 20 questions each per working day. The capacity price is the pay-as-you-go rate of $0.18 per CU hour listed in Azure's retail price list for East US on November 12, 2026, with a 730-hour month.
Per question: (3,000 x 100 + 600 x 400) / 1,000 = 540 CU seconds
Per day: 25 users x 20 questions x 540 = 270,000 CU seconds = 75 CU hours
With caching: (1,000 x 100 + 2,000 x 10 + 600 x 400) / 1,000 = 360 CU seconds
25 x 20 x 360 = 180,000 CU seconds = 50 CU hours per day........
SKU | CU hours per day | Agent share (no cache) | Agent share (cached) | Monthly pay-as-you-go |
F2 | 48 | 156.2% | 104.2% | $262.80 |
F4 | 96 | 78.1% | 52.1% | $525.60 |
F8 | 192 | 39.1% | 26.0% | $1,051.20 |
F64 | 1,536 | 4.9% | 3.3% | $8,409.60 |
........
An F2 cannot carry this workload even with two-thirds of the input cached, and an F4 would leave little room for the warehouse queries the agent triggers, for refreshes and for reports. An F8 handles the questions with about 61% of its daily CU still free before the generated queries are counted. Apportioned to individual questions, 540 CU seconds equals 0.15 CU hours, or about $0.027 at $0.18 per CU hour, but that figure is only an allocation: an F SKU is billed for every hour it runs, whether the agent is used or not. Azure's pricing page states that one- or three-year reservations save about 41% compared with pay-as-you-go.
The same arithmetic gives a capacity ceiling. At 540 CU seconds per question, an F2 supports 320 questions a day if nothing else runs on it, an F4 640 and an F8 1,280. These numbers are illustrative; real token counts vary with instructions, schema size and conversation length, and should be measured in the Capacity Metrics app during a pilot.
··········
COMMON MISREADINGS OF FABRIC DATA AGENT LICENSING.
The first misreading is that a Power BI Pro or Premium Per User workspace is enough. Microsoft's prerequisites name only paid F2+ and P1+ capacities, so a team without one has to buy at least an F2.
The second is that sharing an agent shares the data. The documentation says that configuring or sharing an agent grants no access to its sources, and every answer is filtered by the asking user's permissions, including row-level and column-level security.
The third is that the AI Query charge is the whole cost. The generated SQL, DAX or KQL is billed to its own engine, and on a small SKU those queries can matter as much as the token charge. Similar conversational agents on other platforms, such as Databricks Genie in Slack and Microsoft Teams and the ChatGPT Data agent for BigQuery, Snowflake and Databricks, also separate the agent from the compute it consumes.
··········
FABRIC DATA AGENT COSTS: WHAT DEPENDS ON YOUR CAPACITY AND TENANT.
Microsoft confirms the core facts: the data agent is generally available in the Fabric portal, it needs a paid F2+ or P1+ capacity, it supports up to five sources, it answers in English with at most 25 rows and 25 columns, and it consumes 100, 10 and 400 CU seconds per 1,000 input, cached input and output tokens. Publishing to Microsoft 365 Copilot and connecting to Copilot Studio remain in preview with their own license requirements.
Capacity size and placement are each organization's decision. Token volume grows with instructions, example queries and chat history, the generated queries add engine costs, a region mismatch between agent and data makes queries fail, and tenants outside the EU Data Boundary and the United States must allow cross-geo processing and storage.
Microsoft does not publish a fixed CU cost per question, and the data agent pages do not say whether free-license users on capacities below F64 can chat with an agent built on a semantic model. Admins should check both points in the Capacity Metrics app and with Microsoft before sizing a production rollout, and should decide whether assigning agent users to a Fabric Copilot capacity keeps that usage away from the capacities that serve reports.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
