Skip to main content

EPC Group's Fabric practice builds the semantic model first and the ontology second. Microsoft's ontology generator maps a Power BI semantic model's tables, relationships and DAX measures into entity types, relationship types and metrics, so an uncertified model produces an uncertified ontology that agents then treat as the business. EPC Group certifies before it generates.

Key Facts

  • Two core items: Microsoft delivers Fabric IQ through two core items, the semantic model and the ontology (preview), over data in OneLake.
  • What an ontology is: a shared, machine-understandable representation of the business: entity types, properties, relationships, rules and metrics, bound to data.
  • It is not a data store. It stores definitions and bindings; agents and visuals query the bound sources.
  • Generation: an ontology can be generated from one or more semantic models. Tables become entity types, relationships become relationship types, and DAX measures become metrics.
  • Ownership of the math: the source semantic model remains the owner of each DAX expression.
  • Date: the old ontology experience retires January 31, 2027.
  • EPC Group's position: certify the model, then generate, then finish by hand the five things generation leaves undone.

Last updated by Errin O'Connor, Founder & Chief AI Architect, EPC Group

By Errin O'Connor — Founder & Chief AI Architect, EPC Group · Microsoft Solutions Partner across all six designations · G2 Leader in BI consulting

What is a Fabric IQ ontology?

Microsoft defines the ontology item as a representation of the concepts that matter to an organization, such as customers, products, assets, orders and locations, with their properties, how they relate, and the rules and metrics that give them business meaning. Those definitions are bound to enterprise data so that people, applications and AI agents use the same vocabulary. [2]

The useful word is "bound". An ontology does not copy data. It acts as a virtual context layer over sources that stay where they are: lakehouses, warehouses, eventhouses, SQL databases, mirrored databases and semantic models. [2][4]

What is the difference between an ontology and a semantic model?

Power BI semantic modelFabric IQ ontology (preview)
PurposeCurated analytics: measures, hierarchies, dimensionsBusiness context for operations and agents
UnitTables, columns, relationships, measuresEntity types, properties, relationship types, rules, metrics
Holds dataYes in Import mode, or reads source data through DirectQuery or Direct LakeNo. Definitions and bindings only, unless a graph is materialized
Spans sourcesOne modelMany sources and more than one semantic model
Who reads itReports, Copilot, data agentsAgents, data agents, applications
Owns DAXYesNo. Metrics point back to the source measure
StatusNot in preview (core Power BI item)Preview

Microsoft's description is that the two work together: ontologies can be generated or aligned from semantic models so that terminology and KPIs stay consistent across reports, agents and applications. [1]

A plain way to say it to a board: the semantic model answers "what is the number?" The ontology answers "what is the thing the number is about, and what is it connected to?"

Do you need both?

For reporting and for Copilot answers grounded in certified numbers, the semantic model is enough, and it is where the work starts. EPC Group's Fabric IQ explainer covers that case.

An ontology earns its place when one of three things is true.

  1. A concept spans models. "Customer" lives in sales, support and returns models, and an agent needs one customer.
  2. Agents need relationships, not only totals. Which shipments belong to which order, which assets sit at which site.
  3. Live operations. Time-series and event data from an eventhouse need to attach to the same business entities the reports use.

If none is true yet, an ontology is a preview feature to learn in a sandbox, not a program to fund.

What does generation do, and what does it leave undone?

Microsoft's documentation is specific. Generating an ontology from a semantic model creates: [3]

  • an ontology item in the workspace
  • entity types that match the model's tables
  • properties from the columns, with data bindings
  • relationship types that follow the model's relationships
  • metrics on entity types, from the model's DAX measures

And it leaves five jobs to people: [3]

  1. Bind time-series data. Generation does not create those properties.
  2. Review entity type keys and add any that are missing.
  3. Bind relationship types to data.
  4. Rename entity types and relationships to business names.
  5. Review the whole ontology for completeness.

Job four is where most of the value is. A table called dim_cust_v3 becomes an entity type called dim_cust_v3 unless someone renames it, and an agent will use that name in front of an executive.

Why certify the semantic model first?

Because generation is faithful. It carries over whatever the model says.

  • A number the model only summarizes implicitly has no DAX measure, so no metric is generated for it, and an agent aggregates the raw column its own way.
  • Four models with four definitions of revenue can arrive as four competing metrics unless someone reconciles them.
  • A table with no description becomes an entity type an agent has to guess about.

EPC Group's certification rubric for this is the Groundable Model Standard, described in semantic model certification for AI grounding: explicit measures, complete metadata, a named owner, tested security, sensitivity labels and Prep for AI instructions. The same argument at architecture level is in the semantic model is the new spec.

The Model-First Ontology Sequence

The Model-First Ontology Sequence is EPC Group's build order for Fabric IQ. Each step has an owner.

  1. Certify. The semantic models in scope pass certification. Owner: the certification reviewer.
  2. Generate. Create the ontology from the first certified model. Owner: the data platform lead.
  3. Finish. Complete Microsoft's five manual jobs: time series, keys, relationship bindings, names, review. Owner: the data steward for the domain.
  4. Combine. Add further semantic models through the ontology agent, which reconciles overlapping concepts such as shared customer tables. [3] Owner: the data steward.
  5. Version. Use namespaces to scope definitions by domain, and create named versions as checkpoints you can restore. [2] Apply agent changes as patches, which preserve stable identifiers so downstream queries keep working. [6] Owner: the data platform lead.
  6. Expose. Add the ontology as context to a Fabric data agent (preview), and test answers against known results before any executive sees one. [5] Owner: the analytics product owner.

Skipping step 1 is the common failure. Skipping step 3 is the expensive one.

Who owns an ontology?

Not IT alone. The definitions are business decisions.

DecisionOwner
What counts as a Customer, an Order, an AssetThe business domain lead
The DAX behind each metricThe semantic model owner
Names, keys, relationships and bindingsThe data steward
Tenant settings, capacity, permissions, versionsThe data platform lead
Whether an agent may use itThe AI governance owner

Microsoft requires two tenant settings to create ontology items: Users can create ontology (preview) items, and Users can create Fabric items. Using an ontology with a Fabric data agent or operations agent also needs those agents' tenant settings, and on capacities outside the US and EU Data Boundary the ontology agent is off by default until an admin allows cross-geo processing. Binding a semantic model requires both Read and Build permission on it. [4][6][7] EPC Group scopes the first setting to a named group while the feature is in preview.

How do agents use the ontology?

A Fabric data agent can take an ontology as context. Most semantic configuration, including entity definitions, synonyms, relationships and bindings, lives in the ontology; instructions added in the data agent are used only by that agent and are not written back. [5]

Microsoft's ontology agent can explain an ontology, query the data behind it and propose changes. It chooses the query language by source: KQL for eventhouse tables, SQL for lakehouse tables, DAX for semantic models and GQL for relationship traversals when a graph model exists for the ontology. It shows the query it ran, and when a query returns no rows it says so. [6]

Proposed changes arrive as a patch that preserves stable identifiers and is applied only when the user switches to Act mode. [6] That plan-then-act split is the control to keep: nobody should apply an ontology change they have not read.

What does it cost in capacity?

Ontology requires a workspace on a Fabric capacity. [4] An ontology stores definitions and bindings rather than a copy of the data, so most consumption comes from queries against the bound sources, from the agents, and from any graph a team chooses to materialize. [2] EPC Group sizes that with the method in Fabric capacity FinOps and watches it monthly, as set out in the Fabric F SKU cost model.

What goes wrong?

FailureWhat it looks likePrevention
Generated from an uncertified modelAgents aggregate raw columns their own wayStep 1
Names left as generatedAgents say dim_cust_v3 to an executiveStep 3
Keys not reviewedDuplicate or missing entity instancesStep 3
Two ontologies for one domainTwo definitions of CustomerOne domain, one steward
Changes applied unreadDownstream queries driftPlan mode, then Act
Old experience items left in placeItems stop working after retirementMigrate before January 31, 2027

Quick facts

QuestionAnswer
Is ontology generally available?No. It is a preview item
Can it be generated from a semantic model?Yes, from one or more
What becomes what?Tables become entity types; relationships become relationship types; DAX measures become metrics
Who owns the DAX afterward?The source semantic model
Does an ontology hold data?No. Definitions and bindings only, unless a team opts in to materializing a graph
What is left to do by hand?Time-series bindings, keys, relationship bindings, names, review
What permission is needed to bind a semantic model?Read and Build
What date matters?January 31, 2027: the old ontology experience retires

What changed in 2026?

  • Fabric IQ reached Microsoft 365 Copilot. Generally available since September 2026, Fabric IQ in Microsoft 365 Copilot Chat and Copilot Cowork answers business questions from governed Power BI reports and semantic models. [8] Background: Fabric IQ explained.
  • Generation got richer. The current ontology experience carries DAX measures into the ontology as metrics and includes calculated columns, which the old experience did not. [2][3]
  • An ontology agent arrived in preview. It explains, queries and proposes changes to an ontology, and can combine more than one semantic model. [6]
  • More flexible modeling. Keyless entity types, namespaces, hierarchies, inheritance and named versions. [2]
  • A retirement date was set. Items built on the old experience need a migration plan before January 31, 2027. [2]

How EPC Group delivers this

  1. Model readiness review. EPC Group scores the semantic models in scope against its certification rubric and names which are ready to generate from.
  2. First ontology. One domain, one certified model, generated and then finished by hand with the domain's data steward.
  3. Agent test. A Fabric data agent uses the ontology as context and is tested against questions with known answers.
  4. Operating model. Owners, versioning rules and a change process, written into the client's governance model.

The deliverables are a model readiness report, a finished ontology for one domain, an answer-test log, and an ownership matrix. Engagements are fixed-fee and scoped after a call.

Terms defined

  • Entity type: the reusable definition of a business concept, such as Customer or Shipment.
  • Property: a named fact about an entity, with a data type.
  • Relationship type: a typed link between entity types, such as Customer places Order.
  • Metric: a calculation carried into the ontology from a source DAX measure.
  • Data binding: the mapping from an ontology definition to the data source that supplies it.
  • Namespace: a grouping that keeps definitions organized as the ontology grows.

Best for / Not the right fit for

Best for

  • Organizations with certified Power BI semantic models that are starting to build agents on Fabric.
  • Chief data officers who need one definition of a customer, an order or an asset across domains.
  • Teams on the old ontology experience that must migrate.

Not the right fit for

  • Estates where the semantic models are not yet certified. Start there.
  • Reporting-only use cases with no agent or operational scenario.

Where EPC Group fits

EPC Group has delivered more than 1,500 Power BI deployments and holds all six Microsoft Solutions Partner designations. Ontology work sits inside Microsoft Fabric consulting and Power BI consulting, and starts with a Fabric governance assessment.

The data behind this (sources and verification)

Third-party figures above are attributed to their named sources as of the Last verified date. EPC Group audit figures are directional findings from client engagements.

Frequently asked questions

What is a Fabric IQ ontology?

It is a Fabric item that defines business entities, their properties, relationships, rules and metrics, and binds them to data. EPC Group describes it as the business vocabulary agents read.

Is an ontology the same as a semantic model?

No. A semantic model holds analytical measures and tables; an ontology holds business concepts and bindings across sources. EPC Group builds the model first and generates the ontology from it.

Does an ontology replace Power BI semantic models?

No. Microsoft states that the source semantic model stays the owner of each DAX expression. EPC Group keeps certification effort on the model.

Is Fabric IQ ontology generally available?

No. Microsoft lists it as a preview item. EPC Group runs it in a scoped workspace until that changes.

Can an ontology use more than one semantic model?

Yes, through the ontology agent, which reconciles overlapping concepts. EPC Group assigns one data steward per domain before combining.

What has to be done by hand after generating an ontology?

Time-series bindings, key review, relationship bindings, renaming and a completeness review. EPC Group treats those five as a checklist with a named owner.

When does the old ontology experience retire?

January 31, 2027, according to Microsoft. EPC Group plans migrations for existing items well before that date.

Does an ontology store data?

No. It stores definitions and bindings, and queries run against the bound sources. EPC Group sizes capacity for those queries.

Who should own the ontology?

The business domain owns the definitions and the data steward owns the bindings. EPC Group writes that split into the governance operating model.

What does EPC Group's Fabric IQ work cost?

EPC Group engagements are fixed-fee and scoped after a call with a senior architect. EPC Group does not publish price ranges.

Ready to act on this?

Start with the practice most relevant to your estate, or reach out directly for a senior-architect conversation.

📧 contact@epcgroup.net · 📞 (888) 381-9725

Multiple models. One truth.

AI assistant — not a human