Back to the blog

Why Does a Fabric Data Agent Give Wrong Answers?

Published Nov 18, 2026

The short answer

A Fabric data agent gives wrong answers when it reads your tables without the business meaning behind them, and the wider the data you point it at, the more often it misses. Scope each agent to one subject area, like finance or HR, and train it on your terms. Or ground it in a Fabric IQ ontology, which defines that vocabulary once so every agent that uses it inherits it. Ontology is still in preview, so most teams should start narrow.

Picture the usual first demo. Someone asks the agent for revenue by region. It writes the SQL and the chart looks right. Then a finance lead asks something that depends on how your company defines a booked order, and the agent answers with complete confidence (and the wrong number).

That gap is fixable. If you already have Power BI semantic models your teams trust, you're probably further along than you think. What's left to decide is how wide to scope each agent and whether to wait for ontology to reach general availability. Then there's getting Copilot Studio or Foundry agents onto the same context.

This guide works through those decisions with the live demos from ArchitectNow's Agentic Fabric webinar on 09-09-26, presented by Kevin Harmer, our Chief AI & Data Officer. Every release status below was checked against Microsoft Learn on 11-03-26 and Microsoft's Build announcement, and the full recording is on YouTube.

Why does a Fabric data agent give the wrong answer?

A data agent sees tables and columns. Nothing in them tells it what "active customer" or "booked order" means in your business, so it infers. Those inferences get worse as the data model behind it gets wider and more complex.

Microsoft's name for that missing context layer is Fabric IQ. It provides context on business entities and data, with OneLake as its foundation. In Microsoft IQ it works alongside Work IQ, Foundry IQ and Web IQ (Microsoft Learn: What is Fabric IQ?).

There are a lot of names in this stack (and Microsoft keeps adjusting them), so it helps to think in order instead. Context comes first. You probably have some already in the Power BI semantic models your analysts have relied on for years. A Fabric graph maps how your data connects. An ontology defines your business language and rules, and it can use a graph when a question follows relationships.

Agents come after the context. Data agents answer questions in plain English, and operations agents watch live data and act on it. Planning sits next to both, so budgets and forecasts live in Fabric beside your actuals (Microsoft Learn: What is planning in Fabric?).

Skip the context layer and the agent still answers. It just answers wrong more often, and it can trigger the wrong actions too.

How do you make a Fabric data agent accurate without an ontology?

Keep it narrow, then train it. Point each agent at one subject area, such as one Power BI semantic model or one slice of the warehouse. On a warehouse or lakehouse, teach it your terms with instructions, data source notes and example queries. A semantic model gets most of that through Prep for AI in Power BI.

A Fabric data agent gives people a plain-English way to get answers and charts from your data, and it's generally available (Microsoft Learn: Fabric data agent concepts). One agent can use up to five sources, including lakehouses, warehouses, KQL databases, Power BI semantic models and ontologies, and it writes read-only SQL, DAX or KQL queries for you. It uses the asking person's credentials and permissions, so each answer only reaches data that person is allowed to see. Semantic model sources don't support example queries yet, so use Verified Answers in Prep for AI there. Schema object descriptions, which add business meaning to tables and columns on SQL sources, are in preview and need the preview runtime (Microsoft Learn: Add and configure data sources).

The webinar's rule of thumb was one agent for finance analysis and a separate one for HR. The demo agent covered sales analysis on a subset of a retail data warehouse, and its accuracy came from two things:

  • Scope. Only the tables that matter to sales.
  • Training. Agent instructions, notes on what each data source means, example questions paired with the queries that answer them, and descriptions of the schema.

Asked for total revenue for the 5 best-performing products, as a table and as a line chart across the 12 months of the year, it wrote the SQL and returned both. A follow-up about which product line to prioritize for go-to-market came back with a recommended product line and the product to lead with.

The catch on a warehouse or lakehouse is where that training lives. It sits inside one agent, so a second agent on the same data needs every bit of it again.

What does a Fabric IQ ontology change?

It moves the business meaning out of each agent and into one shared model. An ontology describes your entities and their relationships in business terms and holds your business rules in natural language, so every agent pointed at it starts from the same context and needs far less training of its own.

The webinar's demo ontology covered retail products, stores, sales and freezer installations, and it pulled lakehouse data and real-time eventhouse data into one model. You can start an ontology from a Power BI semantic model you already have and attach the reports that show how the business reads that data.

The webinar showed rules that trigger notifications and actions through Activator built into the ontology. In the new experience, a business rule is a natural-language definition that agents can use to ground their answers. The ontology doesn't run it against data or execute actions, so alerts need an Activator you configure separately or an operations agent (Microsoft Learn: Create and manage business rules in ontology).

The payoff showed when a data agent was set up on that ontology. It didn't need any of the instructions the earlier agent did, because the ontology already carried them.

Microsoft has measured the difference. In an internal benchmark, ontology-grounded agents produced 2.2x more "excellent" responses and a 4.5x win rate in side-by-side comparisons, with 30% fewer tool calls. That's Microsoft's own A/B study, covering 4,000 responses to 400 questions (Microsoft Fabric Updates Blog: Fabric IQ, the shared context layer).

Ontology is still in preview, and it has moved to a new experience; the old one retires 01-31-27 (Microsoft Learn: What is ontology?). When a semantic model is the underlying source, the data agent takes no supplemental context and might not use some of the ontology's context (Microsoft Learn: Use Ontology as context in Fabric data agent). Our advice is to pilot ontology now, in the new experience, and anchor production agents to it once it's generally available.

Data agent without an ontology Data agent on an ontology
Scope One subject area per agent Can widen, because the ontology carries the meaning
Where the business meaning lives Inside each agent on a warehouse or lakehouse; in Prep for AI on a semantic model In the ontology, shared by every agent that uses it
The next agent on the same data Repeat the training on a warehouse or lakehouse Point it at the ontology
Production status Generally available Ontology is in preview

When does a Fabric graph beat SQL joins?

When the question follows relationships several hops deep. A graph describes your lakehouse data as nodes and edges, so one Graph Query Language (GQL) query can walk from a product through its orders to the employees and customers connected to them, where SQL needs a stack of joins.

The webinar demo used retail sales data in a lakehouse, where each order links a product, the employee who processed it, a customer and a vendor. The question: one employee sold one product at the wrong price, so which customers ended up with it? The query added the product, order, employee and customer nodes and filtered on the product number and employee 276. It came back with five customers from a single GQL query.

The graph loads from your lakehouse tables when you save the model, so there's no separate ETL pipeline to set up (Microsoft Learn: How graph in Microsoft Fabric works).

A base graph doesn't describe your data in business terms, though, and that's the job the ontology takes on. Check status before you plan around it. Microsoft's 06-02-26 Build post calls graph generally available, though its Fabric IQ overview still tags it preview. Graph-powered AI reasoning (a graph as a data agent source) is in preview (Microsoft Learn: Graph in Microsoft Fabric overview).

How does a Fabric operations agent act on live data?

You describe a goal in plain English, and the agent turns it into a playbook: a rule that checks live data on a schedule and recommends an action when a condition is met. It can watch an eventhouse directly or through an ontology. The webinar preferred the ontology, but rules on one can't use aggregations or AND conditions yet.

The webinar demo pointed an operations agent at an eventhouse receiving bike counts from bike points across London every few seconds. The goal was one short paragraph: keep at least three bikes at every station, and send a notification naming the station when stock runs low. The agent generated the KQL behind that rule and ran it every 5 minutes. It flagged station 250 with two bikes left in a Teams message, so someone could go rebalance it.

Before you build one, check these against Microsoft's documentation (operations agents, operations agent actions):

  • Operations agents are generally available (Microsoft Learn), but not everywhere: they aren't available in the East US region or in sovereign clouds such as GCC High (limitations).
  • Besides Teams messages, an agent can run a Fabric notebook, pipeline or user-defined function, or trigger a Power Automate flow. Those actions run when the person who receives the recommendation approves it.
  • Each agent gets its own Microsoft Entra identity and acts with its creator's permissions.
  • Trial capacities aren't supported.
  • A state condition such as "is above" signals on every check while the value stays there, and a transition condition such as "crosses above" signals once per crossing.

How do Copilot Studio and Foundry agents use the same Fabric context?

Mostly through MCP, the Model Context Protocol. Fabric exposes its context through MCP servers, so Copilot Studio, Microsoft Foundry and third-party agents can work from the same business vocabulary your Fabric agents use. Choose the route by who should write the answer: the Fabric data agent, or your own agent.

In the webinar, a retail analysis agent in Copilot Studio connected to the demo ontology through MCP and returned the same table and chart the Fabric data agent did. If you're still deciding where your agents should live, our Copilot Studio vs. Microsoft Foundry comparison covers that choice, and how long a Microsoft Foundry agent build takes covers the timeline.

Microsoft's docs name the routes this way today:

  1. Data agent MCP. Use it when the Fabric data agent should work out the answer and the outside agent passes it along. A published data agent becomes an MCP server with a single tool (Microsoft Learn). Microsoft's page carries one warning worth reading before you wire it up: responses can leave Fabric's compliance boundary or region and follow the MCP client's data handling terms.
  2. Fabric IQ Ontology MCP. Use it when your own agent should compose the answer and you have a solid ontology. The more of the business the ontology covers, the better the result. It's in preview (Microsoft Learn).
  3. Fabric IQ MCP. Use it when you have good subject-area semantic models and want your agent to query them directly. It's read-only, respects row-level security, leaves the DAX to your agent, and is generally available (Microsoft Learn).

A Foundry agent can also read an ontology through Foundry IQ as a knowledge source, in preview (Microsoft Learn: Get started with agent integration for ontology).

Which Fabric agent features are generally available today?

Data agents, operations agents, planning and the Fabric IQ MCP server are generally available. Ontology, the Fabric IQ Ontology MCP server and graph-powered reasoning in data agents are in preview. Microsoft's 06-02-26 Build post calls graph generally available, though one Learn table still tags it preview. That's the picture as of 11-03-26.

Capability Status, 11-03-26
Fabric data agent Generally available
Operations agent Generally available
Planning Generally available
Fabric IQ MCP (Power BI reports and semantic models) Generally available
Ontology Preview
Fabric IQ Ontology MCP Preview
Graph as a data agent source Preview
Graph Generally available per Microsoft's Build post; preview tag on the Fabric IQ overview
Data agent as an MCP server Generally available per Microsoft's Build post; no release label on its Learn page

Learn still calls the IQ workload as a whole "IQ (preview)," even though several items inside it are generally available, so check the specific item you plan to run. Expect these rows to change (Microsoft Ignite lands in November).

The rollout path

The webinar closed on a path from the semantic models you have to agents that act, and an attendee question added the sixth step:

  1. Start with the semantic models you already have, and use them to seed your first ontology.
  2. Keep data agents narrow until that ontology is ready and tested on your own questions, then widen them.
  3. Move toward graph and ontology as the way you describe your data, piloting ontology while it's in preview.
  4. Add actions through separately configured Activator rules or operations agents, so a change in the data starts the response.
  5. Connect through MCP so Copilot Studio, Foundry and third-party agents share one business vocabulary.
  6. Put your agents, graphs and ontologies through CI/CD.

The attendee asked about keeping all of this in a single production workspace, and the answer was CI/CD. Change one instruction in an ontology and agent results can shift in ways you didn't plan for, especially through third-party tools, so test those changes before they reach production. More on getting to production: why AI agent pilots stall before they reach the floor.

If you'd rather work through this on your own data, our Fabric Solution in a Day is complimentary: one day with your team and our engineers to prove a data concept on Microsoft Fabric, such as a Fabric IQ ontology or a data agent. You leave with a working solution, in your tenant or ours, plus the plan to scale it. Tell us what you want to prove, and we'll set up a 30-minute call to pick the topic. The recording and slides are on the Agentic Fabric event page.


Sources and references

  1. What is Fabric IQ?, Microsoft Learn, accessed 11-03-26. https://learn.microsoft.com/fabric/iq/overview
  2. What is planning in Fabric?, Microsoft Learn, accessed 11-03-26. https://learn.microsoft.com/fabric/iq/plan/overview
  3. Fabric data agent concepts, Microsoft Learn, accessed 11-03-26. https://learn.microsoft.com/fabric/data-science/concept-data-agent
  4. What is ontology (preview)?, Microsoft Learn, accessed 11-03-26. https://learn.microsoft.com/fabric/iq/ontology/overview
  5. Fabric IQ: The shared context layer for AI agents and real-time applications, Microsoft Fabric Updates Blog, published 06-02-26. https://community.fabric.microsoft.com/blog/fbc_fabricupdatesblogs/fabric-iq-the-shared-context-layer-for-ai-agents-and-real-time-applications/5191678
  6. Graph in Microsoft Fabric overview, Microsoft Learn, accessed 11-03-26. https://learn.microsoft.com/fabric/graph/overview
  7. Create and configure operations agents, Microsoft Learn, accessed 11-03-26. https://learn.microsoft.com/fabric/real-time-intelligence/operations-agent
  8. Operations agent actions, Microsoft Learn, accessed 11-03-26. https://learn.microsoft.com/fabric/real-time-intelligence/operations-agent-actions
  9. Workspace outbound access protection for operations agent, Microsoft Learn, accessed 11-03-26. https://learn.microsoft.com/fabric/security/workspace-outbound-access-protection-operations-agent
  10. Operations agent best practices and limitations, Microsoft Learn, accessed 11-03-26. https://learn.microsoft.com/fabric/real-time-intelligence/operations-agent-limitations
  11. Data agent as Model Context Protocol server, Microsoft Learn, accessed 11-03-26. https://learn.microsoft.com/fabric/data-science/data-agent-mcp-server
  12. Fabric IQ Ontology MCP (preview), Microsoft Learn, accessed 11-03-26. https://learn.microsoft.com/microsoft-copilot-studio/mcp-fabric-iq-ontology
  13. Get started with the Fabric IQ MCP server, Microsoft Learn, accessed 11-03-26. https://learn.microsoft.com/fabric/iq/connectors/fabric-iq-mcp
  14. What's new in Microsoft Fabric, Microsoft Learn, accessed 11-03-26. https://learn.microsoft.com/fabric/fundamentals/whats-new
  15. Create and manage business rules in ontology (preview), Microsoft Learn, accessed 11-04-26. https://learn.microsoft.com/fabric/iq/ontology/how-to-use-rules
  16. How graph in Microsoft Fabric works, Microsoft Learn, accessed 11-04-26. https://learn.microsoft.com/fabric/graph/how-graph-works
  17. Add and configure data sources in Fabric data agent, Microsoft Learn, accessed 11-04-26. https://learn.microsoft.com/fabric/data-science/data-agent-add-datasources
  18. Use Ontology as context in Fabric data agent, Microsoft Learn, accessed 11-04-26. https://learn.microsoft.com/fabric/data-science/data-agent-ontology-sources
  19. Get started with agent integration for ontology (preview), Microsoft Learn, accessed 11-04-26. https://learn.microsoft.com/fabric/iq/ontology/concepts-agent-integration
  20. Agentic Fabric: Your Data Estate Just Got a Workforce, ArchitectNow webinar recording, 09-09-26. https://youtu.be/BJcKWepEN-4

Building something similar?

Talk it through with the engineers who write about it.

Contact us