A regional operations manager asks, “Which region had the lowest on-time delivery rate last quarter?” The question needs an agreed measure, a defined period, the manager's permitted data, and an answer that does not mistake a partial view for the whole company. The chat surface alone settles none of those questions.
Power BI and Microsoft Fabric both let people ask questions like this through native agents, but in different ways. Power BI provides built-in Copilot agents around a report, a published Power BI app, or the content a user can access. Fabric lets a team configure and publish a data agent as its own item, with selected data sources, which other agents can then call.
For where conversational analytics sits among the other Microsoft agent products, see the Microsoft agentic AI stack.
At a glance, there are five deployment routes:
- Power BI agents: Microsoft-managed conversations in a report, an app, or the standalone Power BI experience.
- Fabric data agent: A published Fabric item that answers questions over selected sources with read-only queries.
- Copilot Studio or Foundry agent: A separately built host agent that calls a published Fabric data agent as one of its tools.
- Microsoft 365 Copilot: A published Fabric data agent made available directly in Microsoft 365 Copilot, in preview.
- External agent through MCP: An agent outside those Microsoft hosts that calls either the data agent's question-answering tool or Fabric IQ's lower-level Power BI query tools. MCP, the Model Context Protocol, is a standard interface for an AI client to call tools.
Five routes, three answering capabilities.
Fabric Operations agents monitor data and pipeline runs rather than answer questions; monitoring Fabric data pipelines with the Operations agent covers them.
The Data Behind the Conversation
A Power BI report is a visual presentation of data. Its semantic model defines the reusable measures, relationships, and security rules behind those visuals. If on-time delivery is an approved measure in the model, an agent can use that definition instead of inventing a formula from raw columns.
Fabric contains other analytical sources, including lakehouses, warehouses, and Eventhouse databases for real-time data. Connecting an agent to more sources does not reconcile conflicting definitions. A region name, delivery metric, or reporting period still needs an authoritative meaning in the underlying data, which is the subject of medallion architecture for AI-ready data in Fabric.
Every route raises three separate questions: Which content is in focus? Which data may this caller access? How will the result be checked?
Native Power BI Agents: Report, App and Standalone
Power BI's native agents are built-in Copilot experiences, not custom agent items. A Power BI app is a curated package of reports and related content shared with an audience. The agents' scope changes with the surface:
| Power BI experience | Where the question starts | Content the agent is intended to use |
|---|---|---|
| Report agent | The Copilot pane beside an open report | That report and its underlying semantic model. Generally available. |
| Power BI app agent | A Copilot page inside a published Power BI app | Curated reports and models in that app, subject to the user's access. Preview. |
| Standalone Power BI agent | A full-screen Power BI conversation | Reports and semantic models the user can access; it can also search for accessible apps. Preview. |
All three need a supported capacity and Copilot settings; the standalone experience also needs an administrator to enable its tenant setting.
Marking a model Approved for Copilot improves standalone discovery and removes a low-quality warning. A user may still attach any other item they can access, so the label is not a data-access rule.
Model owners improve answers through Prep data for AI: an AI data schema emphasizes fields, AI instructions clarify vocabulary, and verified answers link a reviewed visual to common phrases. The AI data schema focuses Copilot features that consult model schema; it does not constrain every summary, search, or DAX behavior, and it does not replace row- or object-level security.
To check an answer, a reader can open How Copilot arrived at this to see the fields, measures, and filters used. Authors should test representative questions in the actual report, app, or standalone experience, including with users who have restricted access. Test as role in Power BI does not test Copilot itself.
Native Fabric Data Agent: A Published Analytical Specialist
A Fabric data agent is a configurable item in a Fabric workspace, with draft and published versions. Its author selects up to five supported sources, chooses relevant tables, and adds instructions that help route questions. Typical sources are Power BI semantic models, lakehouses or warehouses, and Eventhouse databases.
The agent generates a read-only query in the language of the selected source: DAX (Data Analysis Expressions) for a semantic model, SQL for a lakehouse or warehouse, and KQL (Kusto Query Language) for Eventhouse. Additional ontology, graph, and search sources have their own preview conditions.
The query language follows the selected source.
Unlike Power BI's built-in agents, this is a specific, reusable specialist that a team publishes and other agents can call. The standard data-agent runtime is generally available; some newer capabilities use a preview runtime or integration.
Its source list and selected schema are focus controls, but sharing the agent does not grant access to those sources. A caller needs the published agent and the underlying data permissions, and row- and column-level restrictions continue to apply.
For semantic-model questions, put field and measure guidance in Power BI's Prep data for AI. Data-agent-level instructions are not passed to the DAX generation tool, and the example-query pairs used for some SQL and KQL sources are not accepted for semantic models.
Authors can inspect the generated query and run steps, then evaluate a test set against expected answers before publishing. A query that passes those checks can still support an incomplete or misleading final sentence.
Copilot Studio and Foundry: Host Agents Around Fabric
Power BI Copilot is not the same thing as Copilot Studio. Copilot Studio is a low-code place to build a separate conversational agent for channels such as Teams. Microsoft Foundry Agent Service is a platform for developers to build and run agents with more code-level control.
In either case, the host agent may call a published Fabric data agent as one tool among others. The data agent answers its configured analytical question; the host handles the broader conversation, channel, and additional tools. That combination is useful when a conversation must mix a data answer with documents, a workflow, or an action.
The host may send Fabric's result to its own model and reshape the final response. Identity differs by host:
- Foundry's built-in Fabric data-agent tool is in preview and uses the signed-in user's delegated identity, so the user needs access to both the agent and its sources.
- Copilot Studio's data-agent tool can use User credentials, so each reader needs source access, or Maker credentials, so readers receive data through the maker's access. Choose User mode when the answer must respect each reader's permissions.
Microsoft 365 Copilot: Direct Publication
A data agent author can also select Publish to Agent Store so that the agent appears in Microsoft 365 Copilot and Teams without a separately built host. This integration is in preview. Each reader needs access to the agent's sources, and row- and column-level security still applies.
Microsoft 365 Copilot has its own orchestrator, which may rephrase the data agent's result. Publishing instructions can ask for the output to be delivered as-is, but they reduce variation rather than guarantee it. Responses may also be processed outside Fabric's compliance boundary under Microsoft 365 terms.
External Agents: Two Different MCP Paths
An external agent can talk to Fabric through MCP, but the available tools do different jobs:
| Tool presented to the external agent | What it supplies | What the external agent still owns |
|---|---|---|
| Published Fabric data-agent MCP endpoint | One natural-language question-answering tool over the data agent's selected sources | The wider conversation, use of the returned answer, and any final-response controls. |
| Fabric IQ MCP | Read-only discovery of Power BI reports and models, metadata inspection, value search, and DAX query execution | Choosing the right model, constructing the query, and explaining the returned data. |
The first path reuses a configured specialist. The second, Fabric IQ MCP, is generally available and gives the external agent more analytical control and more responsibility for query correctness. Agent context through Microsoft IQ explains how Fabric IQ relates to the wider Microsoft IQ family. The Power BI Execute DAX Queries REST API is a lower-level option for a custom application that forms its own DAX.
Identity support differs. Fabric IQ MCP uses delegated user authentication only and enforces the user's existing model permissions. The published data-agent MCP endpoint accepts a user or service-principal (application identity) token, subject to its access requirements. An external host may receive schema, results, or prose outside Fabric's compliance boundary, which makes the connection a data-governance decision.
What Each Route Exposes and How Its Answer Can Be Checked
Testing, inspecting a query, and checking a response before display are different levels of assurance.
| Route | What the agent can reach | How the answer can be checked |
|---|---|---|
| Power BI report | The open report and model, under model security | Test as representative users; inspect fields and filters. |
| Power BI app | Reports and models in the published app, under each user's access | Test the published app with its real audience after changes. |
| Standalone Power BI | Content the user can access; Approved for Copilot filters discovery only | Inspect the chosen report or model, fields, and filters; test restricted users. |
| Fabric data agent | Selected sources and schema, under the caller's source permissions | Inspect generated DAX, SQL, or KQL and run steps; evaluate test questions. |
| Copilot Studio host | The chosen data agent; User mode keeps caller access, Maker mode shares the maker's | Preview and Evaluate test tool choice and answers; the host may rewrite a result. |
| Foundry host | The data-agent tool plus other tools, under the signed-in user's identity | Evaluate test cases and monitor; guardrails check safety, not numbers. |
| Microsoft 365 Copilot | The published agent's sources, under each reader's access | Test as real users; the orchestrator may rephrase the result. |
| External, data-agent MCP | The chosen agent's sources, under the token identity's access | Test the answer; the endpoint returns prose, not a structured query-and-result receipt. |
| External, Fabric IQ MCP or DAX API | Allowlisted reports, models, and measures, under delegated user access | The custom host owns the query, numeric checks, and wording, and can hold an answer until checks pass. |
No route in the table approves every live answer before display. Verified answers come closest for Power BI: for a matching question, Copilot returns an author-reviewed visual, but its text summary is still generated. Verified answers are not returned when Fabric IQ is enabled in report viewing.
When Every Answer Must Pass a Check Before Display
That requirement calls for a custom controlled host, not another instruction telling an agent to be careful. A practical pattern is:
- Use the reader's identity where per-user access matters. Allowlist the approved model, measures, dimensions, and read-only queries; do not use a prompt or AI data schema as the security boundary.
- Obtain a structured result and its provenance through a constrained query tool, such as Fabric IQ MCP or a permitted DAX API call. Retain the model, measure, period, filters, query, and whether the result was complete.
- Have deterministic code compare every stated number, unit, period, and scope with that result.
- Compose the response from checked facts. If evidence is missing or a check fails, withhold the answer, ask a narrower question, or send it for human review.
This is a Reliancy recommended engineering pattern, not a built-in Fabric data-agent setting. Expected-answer evaluations and restricted-user tests still belong before release and after changes.
Show the evidence and the scope of the answer.
Choosing a Route
Native Power BI agents need the least extra infrastructure when the audience already works in reports or an app. A published data agent broadens the sources and can be reused by four routes. Hosts and external agents add channels and control, along with an output and data-movement boundary that the hosting team owns.
Before a pilot, decide:
- which business question and trusted measure the agent must answer
- whether a report, an app, or selected Fabric sources provide the right scope
- where the conversation should live: Power BI, Fabric, a Microsoft host, Microsoft 365 Copilot, or an external agent
- which identity reads the data on each route, including Copilot Studio User or Maker mode
- where answer data may be processed outside Fabric's compliance boundary
- whether tests and inspection are enough, or every answer needs a controlled check before display
For a specific use case, Reliancy can help test the data scope, permissions, and answer evidence across these routes. Discuss a governed conversational analytics pilot.
Sources and Product Documentation
- Copilot for Power BI overview and native agent experiences
- Copilot in Power BI apps
- Standalone Power BI Copilot
- Prepare Power BI data for AI
- AI data schema scope
- Inspect how Copilot answered a data question
- Verified answers and their limitations
- Power BI row-level security and testing limits
- Approved for Copilot and standalone discovery
- Fabric data-agent concepts and runtime
- Create and publish a Fabric data agent
- Fabric data-agent sources and query behavior
- Fabric data-agent sharing and permissions
- Fabric data-agent semantic-model guidance
- Fabric data-agent MCP endpoint
- Fabric IQ MCP consumption tools
- Copilot Studio data-agent tool
- Foundry data-agent tool
- Foundry agent evaluation
- Foundry guardrail intervention points
- Direct Fabric data agent in Microsoft 365 Copilot
- Power BI Execute DAX Queries REST API
- Evaluate a Fabric data agent