Aktiver — semantic AI platform Deterministic · traceable · behind your firewall
Insights / Knowledge graphs

Microsoft puts an ontology under its agents with Fabric IQ

On November 18, at its Ignite conference, Microsoft announced Fabric IQ, a new layer on top of its Fabric data platform. At its centre is an ontology: a model of the entities that matter to a business, how they relate to each other, and what actions can be taken on them, linked to live data in OneLake. Microsoft’s pitch, as VentureBeat reported, is that AI agents need this business context to make reliable decisions, because patterns in raw data are not enough.

It is one of the clearest signals yet that the largest enterprise software vendors now see an explicit semantic model, rather than a bigger prompt or a better embedding, as the missing piece between enterprise data and agents that act on it.

/ 01What Microsoft announced

According to Techzine, Fabric IQ combines an ontology model, semantic models and AI agents, with the stated aim of moving Fabric from a data platform to an “intelligence platform.” The ontology describes a business’s important entities, their relationships and the possible actions, and connects to data in OneLake through Fabric’s engines for analytics, real-time, time-series and geographic data. Techzine reported that the ontology feature was entering private beta, so this is an early release rather than a finished product.

The starting point is familiar to many Microsoft customers. Organisations can bootstrap an ontology from existing Power BI semantic models. Arun Ulag, corporate vice president of Azure Data, told VentureBeat: “We have 20 million semantic models that run in fabric today,” adding that those models had “only used for BI use cases” until now. Fabric IQ extends them with real-time streams and operational rules.

On top of that, Microsoft introduced what it calls operations agents, which Ulag described as agents “that can watch your data for you” and take business actions under human supervision. VentureBeat’s example was an agent that reroutes delivery trucks when congestion appears.

/ 02Why it matters

For years, “ontology” was a word enterprise buyers associated with academic projects and niche vendors. When Microsoft builds one into the foundation of its data platform and explains it to a mainstream audience, the conversation changes. The argument is no longer whether agents need a formal model of the business, but who owns that model, how it is built, and how portable it is.

The announcement also makes a quiet admission. Retrieval over documents and tables gets an agent to plausible answers. It does not tell the agent that a “shipment” belongs to an “order,” that an order has a customer with contractual terms, or that rerouting a truck is only allowed under certain conditions. Those are facts about the business, not about the data, and they have to be written down somewhere the agent can read and obey.

/ 03The knowledge-graph angle

Fabric IQ is a good illustration of a pattern that is spreading: start with a semantic layer that already exists (in this case BI semantic models), promote it to an ontology, bind it to live data, and let agents reason over it. That is the right direction. Three questions decide how far it goes.

  • How rich is the model? A BI semantic model captures measures, dimensions and joins. An ontology that supports reasoning also needs class hierarchies, constraints and clear definitions of what each concept means. Promoting one into the other is useful, but it is a first draft, not the end state.
  • Is it portable? An ontology that lives only inside one vendor’s platform is valuable but locked in. Open standards such as RDF and OWL, and upper ontologies such as BFO, exist so that a model of the business can outlive any single tool and be shared with partners.
  • Can every agent action be traced? The promise of an ontology-backed agent is that its decisions can be explained in terms of the model: which entities, which relationships, which rules. That only holds if the reasoning is recorded against the graph, not left inside a model’s hidden state.
When the biggest data platform vendors start shipping ontologies, the semantic layer has stopped being optional for enterprise agents.

/ 04What to watch

Watch how much modelling depth Fabric IQ exposes as it moves out of early access, whether its ontologies can be exported in open formats, and how its operations agents record why they took an action. For data teams, the practical step is the same whichever platform they use: inventory the semantic models and data dictionaries you already have, and treat them as the seed of a proper ontology rather than as reporting artefacts.

Aktiver’s take — Telecom, media & e-commerce

A major data platform shipping an ontology as its base layer settles the argument about whether agents need one. The next argument is about depth and ownership, and a BI semantic model promoted to an ontology starts thin on both.

Take the truck-rerouting example in an e-commerce fulfilment setting. A useful model says a shipment fulfils an order, the order belongs to a customer with a delivery commitment, and rerouting is permitted only when the new arrival still meets that commitment. With those constraints in the graph, an operations agent can justify a reroute by naming the order, the commitment and the rule it checked, and decline one that breaks the commitment. Support agents grounded in the same catalog and order graph answer “where is my parcel, and why” from facts rather than guesses. Ingestion Hub builds such models from existing tables and documents, lets the people who run fulfilment refine them, and exports OWL or RDF so the model stays usable outside any one vendor.

How we’d model it

ShipmentOrderCustomerReroute actionDelivery commitmentOntology classOneLake tablefulfilsplacedByconstrainedByboundToShipmentOrderCustomerRerouteactionDeliverycommitmentOntologyclassOneLake tablefulfilsplacedByconstrainedByboundTo
  • Shipment —fulfils→ Order
  • Order —placedBy→ Customer
  • Reroute action —constrainedBy→ Delivery commitment
  • Ontology class —boundTo→ OneLake table

See where this applies →

Go deeper

See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.

← All insights Subscribe via RSS → Book a demo →
Keep reading