Setting up AI Agents on top of your Power BI Semantic Models

Data Agents are now generally available in Microsoft Fabric, changing how businesses interact with their data. Instead of relying only on static dashboards, users can now ask questions in natural language and get real insights instantly.

In this guide, I’ll walk you through how to set up an AI Agent on top of your Power BI semantic model — and what actually matters in real-world use.

What are Fabric AI Agents?

A Fabric AI Agent is a chat-based interface powered by AI that connects directly to your data model.

Think of it as:

• A BI Analyst on demand

• A smarter version of Q&A

• A bridge between business users and your semantic model

Prerequisites

Before you start, you will need:

• A Fabric capacity workspace (F2 and above)

• A published semantic model

• Access to the Power BI Admin portal

Step 1: Enable AI Capabilities

Within the Admin portal, you will need to enable the Copilot and Azure OpenAI Service settings under Tenant Settings.

Please note that this also enables:

• Data sent to Azure OpenAI to be processed outside your capacity’s geographic region, compliance boundary, or national cloud instance

• Data sent to Azure OpenAI to be stored outside your capacity’s geographic region, compliance boundary, or national cloud instance

This means that some processing may happen outside your Fabric capacity region.

At this stage, this setting appears to be required. However, as organisations adopt Data Agents more widely, this may change over time.

Step 2: Create Your Agent

To create your agent, go into your Fabric workspace, click New > Data Agent, and then give your agent a name.

A good practice is to give it a role-based name, such as Sales Analyst Agent.

Step 3: Connect to a Semantic Model

Before connecting your agent to a semantic model, make sure the model follows best practices, such as a star schema design.

Once you have connected to the Semantic model, select all the tables you want the Agent to connect to. You can select your measures, fact and dim tables

Step 4: Define Role & Purpose

Within your agent, there is a section for Agent Instructions.

Go to Setup > Agent Instructions and add your prompt there.

This is one of the most important steps, as the quality of the instructions will influence how useful and relevant the agent’s responses are.

Step 5: Testing your agent

Start with a simple prompt such as “Total Sales” and work your way up.

Then try breaking results down by time periods, for example:

• Total Sales by Year

• Total Sales by Quarter

• After that, test across key dimensions such as Sales Manager and Product.

I’ve found that these agents can quickly pick up your style of questioning, which makes testing even more important.

Check for:

Accuracy

Relevance

Consistency

Step 6: Publish your agent

Once you are happy with the model's performance, you can publish your agent

Common Pitfalls

Treating it like magic

• It’s not. It’s only as good as your data model.

Poor instructions

• Generic agents give generic answers.

Weak semantic layer

• If your model isn’t business-ready, the agent won’t be either.