Microsoft is positioning Fabric, its integrated data platform, as the context layer where enterprise AI agents learn how a company operates, whether that's reviewing last quarter's performance or monitoring real-time factory operations. At its FabCon and SQLCon data conference in Barcelona, Spain on Tuesday, the company outlined how Fabric IQ now feeds Microsoft 365 Copilot by default, allows external agents to query it over MCP, and enables Power BI to generate apps from the same definitions. The push responds to what Microsoft Fabric CTO Amir Netz described as agents being "like Drew Barrymore from 50 First Dates" — forgetting everything each time they restart and needing to be told where they are and who they're working for.
The company announced that storage under management in OneLake, Fabric's storage layer, is growing 300% year over year. Power BI users have already created 22 million semantic models, which define how metrics like revenue are calculated and which tables supply the numbers. New integrations this week include a two-way connection with Salesforce Data Cloud 360, mirroring from Google BigQuery reaching general availability, a ClickHouse workload running directly against OneLake, and dbt's Fusion engine arriving in the next couple of weeks. Mirrored security roles, entering public preview in coming weeks, bring permissions along with data so that access roles defined in Snowflake, for example, show up in Fabric with identical members and table permissions. Fabric IQ in Microsoft 365 Copilot Chat and Cowork became generally available Tuesday, answering business questions from Power BI semantic models and reports without consuming additional AI tokens.
According to Microsoft's Yitzhak Kesselman, corporate vice president running Fabric IQ, companies unify their data and run models on it, but more advanced customers with their own evaluations for agents sometimes see unexpected results and realize they need to create context. He met with more than 320 companies last year, with pressure coming from business teams demanding to see "the value of those agents, the before and after." Microsoft's Arun Ulag, executive vice president for Azure Data, argued in the keynote that coding agents succeed because they have access to code repositories, change history, specifications, and tests, but most enterprises outside coding lack anything comparable. The report describes Fabric IQ as combining "unified data from OneLake, trusted metrics from Power BI semantic models, and operational context from ontologies and real-time intelligence."
The system relies on ontologies — formal descriptions of business entities, their relationships, and handling rules — which Netz called "semantic models plus plus." For an airline, these define not just data relationships between pilots and planes but policy relationships like which pilot can fly which plane based on certification or rest hours in the last 24 hours. They also define actions like grounding a plane, redirecting it, or assigning a gate. Creating these ontologies, which must map various company terms to single entities, can be extremely time-consuming, so Microsoft built a tool to generate them automatically. Kesselman emphasized that without streaming data and fresh information, companies end up running language models on data that's hours or days old, recalling two CIOs in Paris telling him "there is no AI without RTI," or real-time intelligence. Power BI Desktop is getting an app-creation experience in preview in coming weeks, where users start from a semantic model, describe an app, and have Copilot generate and publish it to Fabric — apps that can accept inputs, write back data, preserve shared state, and support operational workflows, with Power BI Pro and Premium Per User customers getting a Fabric database of up to 1GB per app at no extra cost. Organizations that master this integration will find their agents can judge current events against historical patterns, but those treating context as an afterthought may discover their AI investments deliver answers that look plausible yet miss the mark. The real test isn't whether agents can access data, but whether the scaffolding around that data reflects how decisions actually get made.