Microsoft recently released Fabric IQ, a new product for building AI agents that are grounded in your organizational data in Microsoft Fabric. In the first part of this 2-part series, I introduced Microsoft Fabric IQ and its semantic layer capabilities: the Ontology and Fabric Graph. The semantic layer is essentially an interface between AI agents and the data that is used to ground the answers of the agent. In this post, I discuss another fundamental aspect of Fabric IQ: the Data Agent and Operations Agent capabilities. I also a discuss the enterprise AI agent architecture in general and how Fabric IQ sits right in the middle of agentic architecture design.
Fabric IQ agents can be viewed as the first agentic layer between the data platform and the applications that consume the data platform. With the Ontology and Graph capabilities, Fabric IQ agents are able interpret the meaning of the data on the data platform and provide context to the raw data. They can be thought as data analysts who:
Fabric IQ recognizes that "AI Agents" are not a one-size-fits-all concept. Broadly speaking, agentic workflows can be categorized into two types:
Fabric IQ addresses these distinct needs with two specialized agent types: Data Agents and Operations Agents. Each agent can be used in various scenarios and by various user groups. In the following sections, I discuss the possible use cases for these agents in more detail.
The Fabric Data Agent is designed for the first scenario: conversational analytics and Q&A. It allows developers to build sophisticated conversational systems that can answer business questions using generative AI, grounded in the organization's data. For example, if you connect the agent to company’s sales and inventory data, you can use the agent to make complex natural language queries like: “What is the sales income from Scandinavia in the last quarter, grouped by country and product category?”.
As we see from the previous example, Data Agent can provide the same information to business users that is traditionally shown in reporting solutions such as Power BI. Building a reporting solution can be a time-consuming task that requires special skills and understanding of the business domain. Furthermore, such reporting solutions have a fixed data model and creating custom reports requires proficiency in report building. With Fabric Data Agents, the reporting solution is not a requirement anymore since business users can converse with the agent directly and the data queries are generated on the fly by the agent.
End users can access Data Agents in various ways. Once an AI developer has published the agent to Fabric, the following user interfaces are available to users:
The following picture shows an example how the Data Agent can be used in Power BI Copilot:
By default, the Data Agent can only be used by those who have access to the agent. The AI developer can grant access to the agent with the standard Microsoft permission management system, backed by Entra ID. It is important to note that Data Agent also respects the data access management system in Microsoft Fabric, so users need access to the underlying data items (Lakehouses, Data Warehouses) as well. This is because the Data Agent always employs the end user’s identity to access the data in Microsoft Fabric.
For AI developers, the Data Agent offers both low-code and code-first experiences to create and interact with Data Agents. The following tools are available for AI developers for these purposes:
It is important to understand that using the Data Agent directly is not the only way to leverage its capabilities. The Data Agent is only the first agentic layer that resides just next to the data platform. We could also build additional agents on top of the Data Agent layer, which use the Data Agent as a tool or source of information. For example, when developing AI agents in Microsoft Foundry, the Fabric Data Agent can be added as a knowledge source for the Foundry agent. The main function of the Data Agent is to interpret and provide data from the data platform, but it could also be used as one of the tools in downstream agents that could have a more general scope and could be able to complete more complex tasks than just data querying.
While Data Agents wait for a user to ask a question, Operations Agents are proactive. They live within the Real-Time Intelligence experience in Fabric and are designed to automate the cycle of Observe - Analyze - Decide - Act for real-time data arriving in Fabric. Real-time data refers to data streams sent from IoT devices or event streams received from other applications, for example.
An Operations Agent is not just a simple alert rule (like "if temperature > 50, send email"). It is a goal-oriented entity. When configuring an Operations Agent, you define:
When you configure these elements, the Operations Agent generates an internal Playbook that is essentially its operating manual. It continuously monitors data streams for patterns that impact its goals.
Crucially, Fabric IQ solves the "trust" problem in agentic AI with a built-in Human-in-the-Loop mechanism. When the agent identifies a condition that requires action, it can send a message to a human operator via Microsoft Teams, for example. This message isn't just a notification; it's a proposal. It includes the context, the reasoning, and a recommended action. The human can approve or reject the action directly in the chat. As trust grows, specific rules can be marked as "autonomous", allowing the agent to act on its own.
The introduction of Fabric IQ marks a shift from a Data Platform to an Intelligence Platform. While a “traditional” Data Platform is primarily designed to collect, store and serve data, an Intelligence Platform understands the meaning and purpose of the data it stores.
An Intelligence Platform goes beyond just processing the data according to pre-determined business logic. It can suggest improvements that align with the business unit’s goals or find hidden trends in the data. It can make conclusions and actions autonomously, but you remain in control of the allowed actions for autonomous actors and shape the goals that the platform is meant to achieve.
Fabric IQ is designed to convert your Data Platform into an Intelligence Platform. If you have already built a Fabric Data Platform for your business, you can use your existing solution as a data source and build the intelligence features on top of it. No need to migrate data or rearrange your data models, the data is already in the correct place in accepted format.
To understand how central the role of Fabric IQ is to the Intelligence Platform, have a look at the following architecture diagram for an agentic AI platform:
The semantic and agentic layers act as interpreters of source data. By querying these layers instead of the raw data sources, end-user applications can formulate their queries dynamically, specifying the desired outcome in natural language instead of writing the query using a specific syntax. This allows for more open-ended applications, such as agentic workflows that don’t have a deterministic output.
Querying the agentic layer instead of raw data has another benefit as well, since it is more robust against changes in the underlying physical data model. Fabric IQ updates the data models in the semantic layer automatically when the physical data model changes in Fabric. Semantic layer abstracts the physical data model so that the applications that use the semantic layer instead of raw data sources don’t need to change their application code when the physical data model changes. This can prevent breaking changes in downstream applications when changing the physical data model in Fabric.
It should also be noted that including a semantic and agentic layer in the Intelligence Platform does not prevent applications from querying the raw data sources directly. If you need reliable output for your application, you can still query the raw data in a traditional way when required and leverage the semantic and agentic layers for more open-ended tasks.
Conclusion
Fabric IQ represents a major leap forward in democratizing agentic AI. By providing the semantic "brain" (Ontology) and the specialized "hands" (Data and Operations Agents), it removes much of the heavy lifting required to build reliable, grounded AI systems.
For organizations, the message is clear: the quality of your AI will depend on the quality of your data modeling. If you want to be ready for the agentic future, start by building your Ontology today.
This concludes our two-part series on Microsoft Fabric IQ. If you missed the first part on the Semantic Layer, read the previous article here.
For more information about the Fabric IQ releases, see these articles on Microsoft blogs: