Fabric IQ: an ontology for your agents, and what it means for your semantic models

Ask an AI agent "which customers are affected by late shipments?" and it has to guess what a customer, a shipment and "late" mean in your tables. This week at Microsoft Ignite, Microsoft introduced Fabric IQ, a new workload in Microsoft Fabric that tries to remove that guesswork by mapping your data to the business concepts it represents.

The hard part of putting agents on business data is rarely the model. It is the meaning. Fabric IQ is in preview, so here is what it is, how it relates to your Power BI semantic models, and when it is worth your time.

What Fabric IQ is, in plain terms

Microsoft describes Fabric IQ as a workload that maps your datasets to the real-world entities they represent, creating a shared semantic structure on top of your data. The centre of it is a new item called ontology (preview).

An ontology is a shared, machine-readable vocabulary of your business. It has four building blocks:

• Entity types, such as Customer, Shipment or Sensor, defined once.
• Properties, the facts about an entity, with declared data types.
• Relationships, typed links like "Customer places Order", with cardinality.
• Data bindings, which connect those definitions to real data in OneLake: lakehouse tables, eventhouse streams and Power BI semantic models.

On top of that sits a graph. Graph in Microsoft Fabric builds an instance graph from your bindings and relationships, so you can traverse connections instead of writing join logic. There is also a natural language query layer (NL2Ontology) that turns questions into structured queries and routes them to the right engine, such as GQL for the graph or KQL for an eventhouse.

The announcement also brings operations agents, a new type of agent in Fabric that uses the ontology to monitor real-time data, spot patterns and take action.

How it relates to Power BI semantic models

This is the question most Power BI teams will ask first: is this a replacement for my semantic model?

No. Microsoft says Fabric IQ is built on Power BI's semantic model technology, and you can generate an ontology directly from an existing semantic model. Generation creates entity types from your tables, properties from your columns and relationship types from your model relationships. You then finish the job by hand: bind time series data, check entity keys, bind relationship types and review the result.

The way I read it: the semantic model stays the home for curated analytics, measures and reporting. The ontology is a wider business vocabulary that spans semantic models, lakehouses and real-time data, and agents use it as context. A well named, well related semantic model gives you a head start.

The preview fine print

The documentation is more useful than the keynote here. A few things stand out:

• Data bindings from a generated ontology depend on the semantic model's storage mode. Import mode models give you entity, property and relationship definitions, but no data bindings. DirectQuery models do not support bindings either.
• Direct Lake is the supported path, and only when the backing lakehouse sits in a workspace with inbound public access enabled. Querying through those bindings works without measures and calculated columns.
• Only managed lakehouse tables are supported, not external tables.
• The graph does not support Delta tables with column mapping enabled, which happens automatically when column names contain spaces or certain special characters.
• The graph does not support the Decimal type, so Decimal columns come back as nulls. Many finance models use Decimal for money.
• You cannot generate an ontology from a semantic model in My workspace.
• New rows upstream only show up after you refresh the ontology item.
• A Fabric admin must enable the ontology tenant settings before anyone can create the item.

None of this is unusual for a first preview. But read the Import mode point twice: many Power BI models are Import mode, so generation will give those teams definitions without data.

When to look at it

After two decades in data, I have seen several "one shared business vocabulary" efforts. Technology was rarely what held them back. Agreeing on what a customer is across sales, finance and operations is the work, and no workspace item does that for you.

So I would look at Fabric IQ when:

• You are building agents that need to reason across domains, not answer questions from one report.
• You have operational or streaming data (sensors, logistics, assets) that you want to connect to business entities.
• Your definitions already live in reasonably clean semantic models or lakehouse tables in Fabric.

I would wait when your main goal is better Power BI reports, when your models are mostly Import mode on capacities you cannot experiment on, or when the business has not yet agreed on its core definitions.

What to do next

A sensible, low-risk way to try it:

1. Ask your Fabric admin to enable the ontology tenant settings in a test tenant or on a test capacity.
2. Pick one small domain with clear entities, for example orders and shipments.
3. Use a Direct Lake semantic model on managed lakehouse tables as the source.
4. Generate the ontology, then review keys, names and relationship bindings by hand.
5. Ask the same five business questions through a data agent with and without the ontology, and compare the answers.

Step five is the point. If the answers get more correct and consistent, it has earned more of your time. If not, you have learned where your definitions are weak.

Takeaway

Fabric IQ puts business meaning next to the data, where agents can use it, and builds on the semantic model work many teams have already done. It is early, and the storage mode limits matter. Start small, use a Direct Lake model, and judge it by the quality of the answers.

Which domain in your organisation has definitions clean enough to try this on first?

Sources

• Microsoft Databases and Microsoft Fabric: Your unified and AI-powered data estate (Azure blog, 18 November 2025)
• What is Fabric IQ? (Microsoft Learn)
• What is ontology (preview)? (Microsoft Learn)
• Generate an ontology (preview) from a semantic model (Microsoft Learn)
• Required tenant settings for ontology (preview) (Microsoft Learn)