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

Data agents have been made generally available to use in fabric environments, changing how businesses interact with their data. Instead of 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 essentially 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 require

• A Fabric capacity workspace (F2 and above)

• A published semantic model

• Access to your Power BI Admin portal

Step 1: Enable AI Capabilities

Within the Admin portal, you will need to enable Copilot and Azure OpenAI Service​s settings within the tenant settings page as below.
*Please note, you are also enabling:

Data sent to Azure OpenAI can be processed outside your capacity's geographic region, compliance boundary, or national cloud instance, and Data sent to Azure OpenAI can be stored outside your capacity's geographic region, compliance boundary, or national cloud instance. This just means that the processing may happen outside your Fabric capacity region.

This setting seems required at this stage, but look out that as organisations adopt Data Agnets there may be changes to this.

Step 2: Create Your Agent

To create your agent, go into your Fabric workspace and click on 'New' -> Data Agent and then give your agent a name (Best practice is to give it a role - like Sales Analyst Agent)

Step 3: Connect to a Semantic Model

Before connecting to a semantic model, make sure the model follows best practices like the Star Schema methodology

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

Step 4: Define Role & Purpose (Most Important Step)

Within your agent, you have a section for Agent Instructions

Go to Setup -> Agent Instructions and add in the prompt

"You are a business intelligence analyst helping users understand performance using this dataset. 
Your goal is to provide clear, accurate, and actionable insights based on the data model

Step 5: Testing your agent

Start with a simple prompt like "Total Sales" and work your way up. Try breaking down the different time periods, like "Total Sales by Year, Quarter, etc." Then work through dimensions like Sales Manager and Product. I find that these agents are great at picking up your traits and style of questioning

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

1. Treating it like magic

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

2. Poor instructions

Generic agents give generic answers.

3. Weak semantic layer

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