AI agents don’t fail for lack of data—they fail for lack of meaning. Microsoft Fabric IQ’s new ontology capability tackles that head‑on by modeling the business concepts, relationships, and rules that live across your estate, then binding them to live data so agents (and people) can ask better questions and take smarter action.
Ontology in Fabric IQ—what it is and why it matters
Ontology (preview) is the semantic backbone of Fabric IQ. You define entity types, their properties, and relationship types, along with rules/constraints. Crucially, you then bind those concepts to real data in OneLake—lakehouse tables, Eventhouse streams, and existing semantic models—so downstream experiences speak a single business language. When you query, predicates get pushed down to native engines (GQL for Graph, KQL for Eventhouse) for scale. Think “Business‑level questions → federated plans over bound sources,” not “stitching three query languages by hand.”
Fabric IQ organizes this into a broader lifecycle: ingest/store (lakehouse, eventhouse, semantic models), model semantics in ontology, visualize/analyze via Fabric Graph and Power BI, and operate/govern with versioning, validation, and health. It’s a first‑class item in the IQ workload.
Two design moves make ontology practical:
- Generate from what you already have. You can seed an ontology directly from a Power BI semantic model, then enrich it with time‑series or operational streams.
- Preview as a graph and in tables, then ask in natural language. The preview experience shows instance data and a graph view, and Fabric data agents let you query in plain English over ontology‑bound sources.
Where ontology sits in the Fabric stack
Graph in Microsoft Fabric (preview) brings labeled property graph modeling and GQL standards support; ontology can automatically build a navigable graph and push graph predicates to GQL. This gives you visual exploration and pattern queries grounded in your business vocabulary.
On the conversational side, Fabric data agents combine up to five sources—including ontologies, lakehouses, warehouses, KQL databases, and semantic models—so users and apps can ask governed, natural‑language questions across the same semantic fabric.
And because semantics without stewardship is fragile, Microsoft Purview’s Unified Catalog and Purview Hub for Fabric provide the governance layer: catalog discovery, sensitivity labels, DLP for semantic models, auditing, and active glossary terms/policies that travel with data products.
Status note: Ontology and Graph are in public preview as of November 19, 2025; experiences and limits can change. Check the Fabric “What’s new” and IQ docs for current details.
How the ontology works (capabilities that matter)
Modeling & constraints. You define entity/relationship types, properties, and constraints. This isn’t just a glossary—constraints shape how instances relate.
Binding to data. Bindings declare identity keys, property mappings, and cross‑source relationship keys so ontology instances materialize from lakehouse tables, eventhouse streams, and/or semantic models. You get a coherent business view over heterogeneous engines—without ETL sprawl.
Querying. Ontology queries start from business entities and traverse relationships, with pushdown to GQL and KQL where appropriate. Data agents layer natural language on top; Graph offers visual pattern‑finding.
Operate & govern. Fabric IQ lets you version and validate ontology assets and monitor their health. Pair this with Purview’s catalog, labeling, DLP, and audit for end‑to‑end control.
Playbook examples you can steal
Below are three sector‑specific patterns you can adapt. Each follows the same rhythm: model the concepts, bind to sources, explore/ask, then let an agent act with policy guardrails.
1) Financial services: card‑fraud and exposure management
Model. Customer, Account, Card, Transaction, Merchant, Device, Alert. Relationships: Customer‑owns‑Account, Account‑has‑Card, Card‑usedAt‑Merchant, Transaction‑originatesFrom‑Device. Bind cards/transactions from the warehouse/lakehouse; stream device and merchant risk signals via Eventhouse.
Ask. “Show customers with a sudden 3× increase in cross‑border transactions this hour at merchants connected (within 2 hops) to prior fraud alerts.” Ontology pushes time filters to KQL for stream windows, relationship patterns to GQL for graph neighborhoods.
Act. A data agent assembles context and proposes limits or step‑up authentication. An Operations Agent (announced as part of Fabric IQ) can run with human‑in‑the‑loop approval for automatic holds above a threshold.
2) Oil & gas: production optimization across wells and equipment
Model. Well, Pad, Compressor, Sensor, WorkOrder, SafetyIncident, Vendor. Relationships: Well‑locatedOn‑Pad, Compressor‑services‑Well, Sensor‑monitors‑Compressor, WorkOrder‑targets‑Asset. Bind asset master data from the lakehouse, telemetry from Eventhouse, and maintenance history from a warehouse/mirrored system.
Ask. “Which wells will fall below target throughput in the next 24 hours given rising inlet temperature at upstream compressors and recent vendor MTBF?” The ontology query traverses asset relationships while the time‑series predicate pushes to KQL. A graph neighborhood shows the upstream chain in one view.
Act. An Operations Agent drafts a rebalancing plan and a work order sequence, respecting safety constraints and vendor SLAs for dispatch—escalating to a human if risk exceeds policy.
3) K‑12 education: attendance, intervention, and support
Model. Student, Enrollment, Course, Teacher, AttendanceEvent, Assessment, SupportPlan, Guardian. Relationships: Student‑enrolledIn‑Course, Course‑taughtBy‑Teacher, Student‑has‑Guardian, Student‑has‑SupportPlan. Bind SIS tables in the lakehouse; attendance events stream through Eventhouse.
Ask. “Identify students with >10% unexcused absences over the last 30 days who lack a current support plan and are enrolled in >3 courses flagged as high‑risk.” Ontology unifies “student” across systems; the time window pushes to KQL; the traversal checks support plan relationships.
Act. A data agent prepares counselor briefs; an Operations Agent can auto‑schedule outreach within district policy. (Human approval gates remain configurable.)
Governance that scales with meaning
Semantics without governance can backfire. In Fabric, Purview’s Unified Catalog inventories Fabric items, applies sensitivity labels, supports DLP for semantic models, and logs user actions via Audit; the Purview Hub surfaces this right inside Fabric. Glossary terms in the Unified Catalog aren’t just labels—they can carry active policies applied to data products for consistent right‑use. Together, this keeps ontology‑driven experiences safe, discoverable, and measurable.
Getting started (now, while it’s in preview)
- Seed an ontology from an existing Power BI semantic model; then bind lakehouse tables and Eventhouse streams for operational context.
- Preview and validate: use ontology’s instance/graph views and run a couple of natural‑language questions through a Fabric data agent.
- Close the loop: publish a governed, minimal end‑to‑end scenario (e.g., fraud alert triage or attendance triage) with Purview labels and audit enabled.
Tip: Use Graph in Fabric where relationship‑heavy questions dominate; keep fast aggregates in your Power BI model. Ontology gives you the language to move between them consistently.
Limits and expectations
This is preview software. Feature surface, performance characteristics, and licensing details are evolving. Graph and Ontology are actively shipping updates; Operations Agent is newly announced and still rolling out. Validate workloads and revisit limits before committing critical runbooks.
Bottom line
Fabric IQ’s ontology turns data into shared meaning—the missing piece for agents that reason, explain, and act. By modeling real‑world concepts, binding them to live sources, and governing the whole lifecycle, teams can move from dashboards about the past to decisions about the next best action. Start small: one domain, one high‑value decision loop, and the governance you’d be proud to scale.
EduDataSci - Educating the world about data and leadership 