AI Agents

An AI agent answers questions about your Power BI data in plain language. You build the agent in Microsoft Foundry and set it up there, and then connect it to the portal under AI Agents in the user menu. The portal hands the agent the semantic model it should work with and a way to query it, so the agent looks up the actual figures before it answers. A connected agent can be used in two ways:

• beside a report, as a chat panel next to a Power BI report in the portal,
• in Embedsy Studio, in the AI Agent component of an app and in the DAX query editor.

Adding agents and attaching them to semantic models is for admins.

Create the agent in Foundry

The portal doesn't create agents: it connects to one that already exists in a project in Microsoft Foundry. There:

1. In your Foundry project, create an agent. Give it a name, choose the model deployment it runs on (for example gpt-5-mini), and add the instructions, tools and guardrails it should have. Save it. Foundry keeps every saved change as a new version, and the portal always runs the latest one.
   
2. On the project's Project details page, copy the Project endpoint. It looks like https://<resource>.services.ai.azure.com/api/projects/<project>; its last part is the project's name.
   
3. Open the model deployment the agent runs on, go to its Details tab and copy the Key. Take only the key from this tab: the Endpoint next to it belongs to the model, not to the project.
   

Add the agent to the portal

Open the user menu at the bottom of the portal navigation and choose AI Agents. Only admins see it. The page lists the agents already connected.

Click Add from Foundry. The Add a Foundry agent dialog asks for:

• Foundry project - the project's name, shown as Name on the Project details page and as the last part of the project endpoint.
• Project endpoint - the endpoint you copied from Foundry.
• API key - the key you copied. It is stored encrypted and never shown again.
• Foundry agent - open the list to load the project's agents and pick one. Each entry shows the agent's model, version and the start of its instructions. If the list is empty, the endpoint or key is wrong, or the project has no saved agent.
• Display name - optional, the name people see in the portal. It defaults to the agent's name.

Click Test connection. When Foundry answers, the button turns green and reads Connection works. Then click Add.

Each card on the page shows the agent's project, its Foundry agent and its endpoint, with icons to Edit connection, Share and Remove model. A Foundry agent can be added only once per portal.

Changing or removing a connection

• Edit connection opens the same dialog. Leave the API key blank to keep the stored key; if you change the project endpoint, you have to enter the key again.
  
• Remove model takes the agent out of the portal; it stays untouched in Foundry. Removing is refused while a semantic model or a Studio app still uses the agent, and the message names them, so pick another agent there first.

Two ways to use an agent

Once an agent is added, it can answer questions beside a report and in Embedsy Studio. Each is switched on in its own way.

Beside a report

An agent can sit next to a Power BI report in the portal and answer questions about the report's data. For that, three things must be in place:

1. On the Semantic Models screen, an admin ticks Open to LLMs on the report's semantic model.
2. On the same card, the admin picks the agent under AI agent. It then serves every report built on that model.
3. The viewer's role has Can use AI agents ticked, on the Roles tab or on the Entra role. New roles start without it.

The agent then appears as a narrow bar with its name on the right of the report. Click it to open the chat, drag the panel's edge to widen it, and use the arrow to collapse it again. The conversation starts afresh on another report.

The agent reads the semantic model behind the report, not the report page: it doesn't see the visuals, or the filters and slicers the viewer has set. Name them in the question ("for crude oil in 2018"), or the answer may cover all the data.

In Embedsy Studio

In Studio, the person building an app or writing a query picks the agent and the semantic model it works with:

• an AI Agent component puts a chat with the agent into an app, for the app's viewers;
• the AI assistant of the DAX query editor uses the agent to help write DAX queries.

The prompt library

While configuring an AI Agent component you can save any question and the answer it produced. Saved prompts go into one library for the whole portal, not tied to an agent or app.

Clicking the star on a prompt endorses it. Endorsed prompts are offered to viewers of every published app with an AI Agent component, as suggested questions they can click to load into the chat box; prompts that are not endorsed are only shown while building. Endorsing a handful of well-phrased questions is the simplest way to show viewers what the agent is good at. The suggestions can be switched off per component - see AI Agent.

What the portal adds to each question

The agent runs in Foundry with the model, instructions, tools and guardrails you gave it there, and the portal leaves it unchanged. With each question, the portal adds a description of the semantic model (its tables, columns and measures, read live from Power BI) and lets the agent run read-only DAX queries against that model: the agent writes a query, the portal runs it, and the agent answers from the rows that come back. The queries it ran appear in the answer's Thoughts section. Each question allows a few rounds of queries; if the agent uses them up, for example while correcting a query that failed, it answers from what it has or asks whether to go on.

An agent doesn't carry a semantic model of its own: beside a report it uses the model the report is built on, and in Studio the one chosen there, so the same agent can serve several models. Only models marked Open to LLMs on the Semantic Models screen can be used. Without a model, the agent has no access to your data.

Troubleshooting