Fabric IQ, decoded

Once Copilot schedules a meeting based on what it read in your semantic model, a wrong measure description produces a wrong action, not just a wrong number.

Microsoft published a demo of this in June 2026. A Finance report: revenue stable, margins under pressure. Someone asks Copilot Cowork to summarize the trends. Fabric IQ reads the semantic model, finds the margin signal, pulls in relevant email and Teams threads, drafts a leadership follow-up, schedules a cost review, and starts drafting a PowerPoint — all from a single prompt by someone who never opened a data tool.

Fabric IQ is the layer that makes this possible. It connects your semantic model to M365 workflows so agents can move from insight to action without a human routing them.

Fabric IQ is one of four workloads inside Microsoft IQ, the enterprise intelligence layer Microsoft is building into the stack. The others are Work IQ (Microsoft 365 signals: emails, Teams, meetings), Foundry IQ (enterprise documents and policies), and Web IQ. Each is standalone; Copilot Cowork combines them — Fabric IQ read the Finance report data, Work IQ read the Teams threads, and together they routed from insight to action.

Fabric IQ itself has three layers. OneLake is the data foundation. Power BI semantic models sit above it as the business intelligence layer: curated measures, hierarchies, and dimensions that agents reason over. The ontology, still in preview, is the operational intelligence layer: it defines business entities, relationships, rules, and actions so agents can do more than answer questions.

The IQ workload in Fabric is in preview. The semantic model is GA. The ontology and operations agents are preview.

Patrick LeBlanc made this point in a recent Guy in a Cube video. A report with a wrong model produces a wrong number. An agent with a wrong model schedules the wrong meeting. "That is a different level of risk."

Fabric IQ reasons over what your model describes: measure names, descriptions, table context. Fabric IQ falls back to surrounding context to interpret a measure named Rev_Adj_YTD with no description. When that interpretation shapes a leadership deck, the cost of ambiguity rises.

Describing your measures used to be a week of documentation work. Three built-in functions change this: INFO.VIEW.MEASURES(), INFO.VIEW.COLUMNS(), and INFO.VIEW.RELATIONSHIPS(). Each returns a dynamic metadata table from your model. Copy the measure table from your TMDL view, paste it into any AI with a prompt to write descriptions, apply back. Maven Analytics published a walkthrough of the full workflow in June — link in WORTH READING.

The semantic model now drives both your reports and your agents. The bar for accuracy is the same. The consequences for missing it are not.

Six connectors are moving from embedded ODBC to Apache Arrow Database Connectivity: Databricks, Dremio, Google BigQuery, Snowflake, Spark, and Impala. Hive is the seventh affected connector — deprecated entirely with no ADBC replacement. Legacy ODBC stops executing in the Power BI Service in September 2026; Desktop and Gateway follow in March 2027. Opt in per connection by adding Implementation="2.0" to the connection string, or enable it at the workspace or tenant level. Miss September and you need an on-premises Gateway to keep things running.

Operations Agents are in preview in Fabric Real-Time Intelligence. They consume the ontology layer for business context (entities, relationships, actions, rules) and take action based on what they find. They understand operational meaning, not just query syntax. The semantic model feeds the ontology, which feeds the agent.

The Fabric MCP Server is in public preview. MCP-compatible agents and tools — GitHub Copilot in VS Code, Claude Code, others — can connect to OneLake data directly through it: browse workspaces, read files, discover table schemas. Requires Azure CLI installed and authenticated. No custom connectors needed.

The Power BI report authoring skill, part of Microsoft's skills-for-fabric repo, lets a GitHub Copilot CLI agent create and edit PBIR reports from natural language prompts. Guy in a Cube tested it.

Basic prompt ("create an overview page and a store details page"): functional report, two pages with similar content, took 10 minutes. Not visually impressive.

Design inspiration prompt (drop a screenshot or reference report into the PBIP folder, prompt "make it look like the inspo file"): it copied the rounded corners, colors, and font sizes. That part worked.

The design inspiration approach is the useful one: add your brand guidelines or a reference design to the PBIP folder, prompt once, get a styled starting structure. The output is PBIP/TMDL, diffable and reviewable in Git. Use it to get a starting point, not a finished report.

Status: Preview. Requires GitHub Copilot CLI and the Fabric Skills package installed.

The semantic model now drives your reports and your agents. The bar for accuracy is the same. The consequences for missing it are not.

One sentence forward CTA - specific to this issue’s topic.