In a previous post, I argued that ontologies, semantic models, and dimensional models are not competing alternatives. They are three distinct layers of the same thing, each a derived, increasingly constrained projection of the one above. The ontology captures what your business fundamentally is. The semantic model translates that into something analysts can query. The dimensional model flattens it further into something a report can render.
The hierarchy runs downward: ontology → semantic model → dimensional model.
Most of us in the Microsoft ecosystem have been building in the opposite direction. We start with tables, add a semantic model on top, layer in measures and relationships, and call it a single source of truth. That work represents years of accumulated institutional knowledge: what a customer is, what a sale means, what counts as revenue.
The problem is that all of that knowledge is locked inside a Power BI dataset. It lives in DAX. It exists to serve visuals. It cannot explain itself to an AI agent, cannot be shared across systems without rebuilding from scratch, and cannot evolve without a developer.
That is precisely the problem Fabric IQ is designed to solve.
What Microsoft Actually Announced #
Microsoft introduced Fabric IQ at Ignite in November 2025 [2]. The pitch, delivered by Amir Netz, was ambitious: elevate Fabric from a data platform to an intelligence platform. The language leaned hard into “semantic intelligence” and “agentic AI,” which is reasonable cause for skepticism. We have heard that kind of language before.
Here is what is underneath it.
Fabric IQ introduces a new workload within Microsoft Fabric. At its centre sits a new artefact type called the Ontology, currently in public preview [4]. It also includes integration with the existing Data Agent, a new Graph capability for traversing entity relationships visually and via GQL queries, and a forthcoming Operations Agent designed for autonomous real-time monitoring. Semantic models, which already existed, are now formally part of the IQ workload alongside ontologies.
The core idea is that Fabric has done a good job solving where data lives. OneLake is a real answer to data sprawl. What it has not solved is what that data means. An ontology in Fabric IQ is the structure that answers that question in a way that both humans and AI agents can read [1].
The Ontology in Practice #
If you have built a Power BI semantic model, the concepts will be familiar. Tables become entity types. Columns become properties. Relationships stay relationships. The vocabulary is similar enough that the mental model transfers.
But there is one thing that is fundamentally different.
A semantic model is technical. It exists to serve queries. It is optimised for DAX, for visual consumption, for a specific reporting context. A Customer table in your semantic model might join to five different fact tables, but what it means to be a customer, what rules govern that definition, what actions a system is allowed to take on behalf of a customer, none of that lives in the semantic model.
An ontology captures the meaning. A Customer entity type in an ontology can reference columns from multiple data sources, carry business rules as constraints, describe relationships to other entities, and expose permitted actions. It is designed to be read by an AI agent that has no prior context about your business, and give that agent enough grounding to reason correctly.
Constellation Research analyst Michael Ni put it plainly at Ignite: “Ontologies don’t build themselves.” [7] He is right. There is real upfront work here. The democratisation pitch, that business experts can build ontologies themselves using no-code tools without waiting on engineers, is plausible in theory. In practice, the quality of your ontology will reflect the quality of your shared understanding of the business. If your sales team and your finance team currently define “revenue” differently, that dispute does not disappear because you have a new artefact type to argue about. The tool does not resolve the organisational problem. It just makes the problem more visible.
That is actually useful. Visibility is the first step.
The Bootstrapping Question #
Microsoft has made it easy to generate an ontology from an existing semantic model [5]. You point it at a Power BI dataset, and it produces entity types from your tables, properties from your columns, and relationship types from your model’s relationships. For straightforward domains, this gets you most of the way there. Customers, products, orders, transactions.
I want to be honest about the limitation, because it matters architecturally [6]. If you generate an ontology upward from a semantic model, you are inheriting the constraints of that semantic model. You are not modelling what your business fundamentally is. You are modelling what your BI layer currently implies. Those are different things.
The semantic model was built to answer questions that existed when someone set it up, probably two or three years ago. It reflects the reports needed then, the data available then, the definitions people agreed on then. Generating an ontology from it is smart and practical. It reuses real work. But it is a starting point, not a destination.
The previous post ended with a challenge: if you want an ontology that actually represents your business, the direction should be top-down, not bottom-up. Model what the business is, then derive the semantic model from it, then the dimensional model from that. Fabric IQ supports this too. The upward bootstrap is the pragmatic on-ramp for organisations with an existing Fabric investment. Long-term, the hierarchy should run the other way.
Where This Fits #
Fabric IQ does not exist in isolation. Microsoft has framed it as part of Microsoft IQ, which also includes Work IQ (how people work, from M365), Foundry IQ (policies and authoritative documents), and Web IQ (the open web) [1]. The idea is that AI agents grounded across all four layers have a coherent picture of the organisation to reason against. It is an interesting architecture. It is also very early.
Ontology is in public preview. The tooling has real rough edges. The data binding story behaves differently depending on whether your semantic model runs in Import mode, Direct Lake mode, or DirectQuery mode, and some combinations do not give you a fully queryable knowledge graph. If your data sits in a workspace with public inbound access disabled, which is the right security posture for most production environments, you hit binding limitations Microsoft has not fully resolved yet.
At FabCon in Atlanta in March 2026, Microsoft added Planning to the IQ workload, bringing budgets, forecasts, and scenario modelling directly into Fabric’s semantic layer [8]. Actuals and plans in the same place, grounded in the same definitions, accessible to the same agents. That is a significant directional statement. Microsoft is positioning Fabric IQ not just as context for AI, but as the foundation for how organisations make decisions.
What This Means for the Work We Do #
Here is what I keep coming back to.
The problem Fabric IQ is trying to solve is real. AI agents that reason against raw tables and column names make mistakes because they lack context. The context they need is business meaning, and business meaning has historically been scattered across semantic models, data dictionaries that nobody reads, wikis nobody updates, and the heads of people who have been around long enough to remember the decision.
An ontology is a serious attempt to codify that context in a durable, machine-readable form. The concept is not new. What is new is that Microsoft has built it into the platform that a very large number of organisations are already running their data work on.
Whether it succeeds will depend less on the technology and more on whether organisations are willing to do the hard work of agreeing on definitions. The community has said it clearly and I agree: the agent is only as reliable as the context it has to work with [3]. If your semantic model has ambiguous column names and undocumented measures, an ontology built on top of it inherits all of that. Garbage in, hallucination out.
Get the meaning right first. The agents will follow.
References #
[1] Microsoft Learn. What is Fabric IQ? November 2025. https://learn.microsoft.com/en-us/fabric/iq/overview
[2] Yitzhak Kesselman, Microsoft. From Data Platform to Intelligence Platform: Introducing Microsoft Fabric IQ. Microsoft Fabric Blog, November 2025. https://blog.fabric.microsoft.com/en-us/blog/from-data-platform-to-intelligence-platform-introducing-microsoft-fabric-iq
[3] Chafia Aouissi, Microsoft. Fabric IQ: The Semantic Foundation for Enterprise AI. Microsoft Fabric Community Blog, November 2025. https://community.fabric.microsoft.com/t5/Fabric-Updates-Blog/Fabric-IQ-The-Semantic-Foundation-for-Enterprise-AI/ba-p/5172473
[4] Microsoft Learn. What is Ontology (Preview)? April 2026. https://learn.microsoft.com/en-us/fabric/iq/ontology/overview
[5] Microsoft Learn. Generate an Ontology from a Semantic Model. 2026. https://learn.microsoft.com/en-us/fabric/iq/ontology/concepts-generate
[6] Michael Ridland, Team 400. Generating a Fabric IQ Ontology from a Semantic Model: What It Does and Where It Falls Short. May 2026. https://team400.ai/blog/2026-05-fabric-iq-ontology-from-semantic-model
[7] Anirban Roy, InfoWorld. Microsoft Fabric IQ Adds ‘Semantic Intelligence’ Layer to Fabric. November 2025. https://www.infoworld.com/article/4093181/microsoft-fabric-iq-adds-semantic-intelligence-layer-to-fabric.html
[8] Microsoft Fabric Blog. Introducing Planning in Microsoft Fabric IQ: From Historical Data to Forecasting the Future. March 2026. https://blog.fabric.microsoft.com/en-us/blog/introducing-planning-in-microsoft-fabric-iq-from-historical-data-to-forecasting-the-future/
Photo by ChatGPT