Use Generative AI in Microsoft Copilot Studio

Scenario

In this lab, you will prepare your agent for production use by configuring grounding, fallback behavior, and safety controls while keeping generative AI enabled.

In this exercise, you will:

Configure your agent to use generative AI only within approved knowledge sources
Guide your agent to respond predictably and safely when it cannot answer a question

This exercise will take approximately 20 minutes to complete.

What you will learn

How to constrain generative AI using grounding
How to configure system topics for safe fallback behavior

Prerequisites

Must have completed Lab: Create agent flows

Detailed steps

Exercise 1 - Configure authentication

Task - Configure authentication

If it’s not still open, go to the Microsoft Copilot Studio portal https://copilotstudio.microsoft.com and ensure you are in the appropriate environment.
Open the Real Estate Booking Service agent.
Select Settings in the upper-right of Real Estate Booking Service.
Select the Security tab.
Select Authentication.
Select Authenticate with Microsoft.
Select Save.
Select Save in the confirmation window.
Select the X in the upper-right to close out of the Settings.
Select Publish.

Exercise 2 - Ground generative answers to approved data

In this exercise, you will configure generative answers so the agent can only respond using approved knowledge.

Task 2.1 - Restrict generative answers in Conversational boosting

Select Topics
Filter by System topics.
Open the Conversational boosting topic.
Select the Create generative answers node.
Select Edit for Data sources.
Select Search only selected sources.
Select Add knowledge.
Select the Real Estate Property Dataverse table and Add to agent.
Check both the website and the Real Estate Property table as the knowledge sources.
Scroll down and disable Allow the AI to use its own general knowledge.
Select Save.

The agent can no longer answer open-ended questions using general model knowledge.

Exercise 3 - Replace default fallback with a grounded fallback

In this exercise, you will reconfigure the system fallback behavior.

Task 3.1 - Use generative answers in System fallback topic

Select the Topics tab and select the System filter.
Select the Fallback topic.
Select the ellipses (…) menu in the Message node and select Delete.
Select the the + icon under the Condition node, select Advanced > Generative answers.
In the Input field, select the System tab and select Activity.Text.
Select Edit under Data sources.
Select Search only selected sources.
Add the website and the Real Estate Property Dataverse table.
Scroll down and disable Allow the AI to use its own general knowledge.
Select the Customize checkbox under Content moderation level and select Medium.
Select Save.

When an agent cannot answer a question, it will respond safely using grounded data instead of guessing.

Task 3.2 - Add a controlled fallback response

In the Fallback topic, add a Send a message node after the Generative answers node.
Enter the following message: Sorry, I don’t have enough information to help with that request. I can assist with booking real estate showings or questions related to available properties.
Select Save.

The agent now communicates scope and limitations clearly instead of generating ambiguous responses. Note that the Fallback topic only triggers when the agent cannot confidently select any topic of generative answer path. The Fallback topic is only for unknown intent.

Exercise 4 - Validate production behavior through action

In this exercise, you will ask your agent questions that are both inside and outside of its scope of knowledge.

Task 4.1 - Trigger a supported scenario

Open the Test panel.
Start a new test session.
Ask a question that is supported by your configured knowledge, such as Is there a property on Oak Lane?

The agent should respond using grounded, approved data.

Task 3.2 - Trigger an unsupported scenario

In the same test session, ask a question outside the agent’s scope, such as What's for dinner?

The agent should avoid guessing at an answer.

Summary

In this lab, you actively configured generative AI behavior for production readiness by:

Grounding generative answers to approved data
Replacing default fallback behavior
Enforcing safe, scoped responses

Challenge - Add knowledge from a file

For an added challenge, add a file as an additional knowledge source for your agent. Ensure that it is also added for Conversational boosting.

To download the file: open a new window and navigate to https://download.microsoft.com/documents/customerevidence/Files/4000007499/SummitRealtyCaseStudy.docx