As AI technology advances, it offers organizations multiple opportunities to provide more and better ways to improve productivity and guide employees. One way to achieve this is through generative AI. Copilot Studio allows you to leverage generative responses in several ways. Once in Copilot Studio, generative AI features can be accessed by selecting Generative AI in the Settings menu.

With Copilot Studio, you can use generative AI to:

• Use generative responses as a fallback: Allows you to use generative responses if your agent cannot find a relevant answer to the user’s question.
• Insert generative responses into topics: Enables you to integrate generative AI into your topics using the generative response node.
• Use Copilot to create agents and topics: Copilot allows you to provide a brief description of the agent or topic you want to create, and it will build it for you.

Use generative responses as a fallback

Previously, if an agent could not determine a user’s intent, it would ask them to rephrase their question. If the agent could not identify a topic after two attempts, it would escalate to a human agent.

With generative responses, Microsoft Copilot Studio enables your agent to find and present information from multiple sources, internal or external, without creating topics. This allows you to use generative responses as primary information sources or as a fallback when created topics cannot answer a user’s query. As a result, this significantly reduces the time required to create and deploy a functional agent, eliminating the need to manually write multiple topics that might not cover all customer questions.

Generative responses use knowledge sources as “anchor data.” Anchor data helps provide more context. For example, when you use an internal data source as a knowledge source, the source is based on your data, so the information presented to the user will be more relevant than generic information from a public site. Several knowledge sources are currently available, and more are continuously being added.

This article helps you get started using generative responses as a fallback topic when a user’s intent cannot be handled by the agent’s existing topics.

Generative responses can use these sources:

External resources:

• Bing Web Search — requires no external configuration
• Bing Custom Search — requires external configuration

Internal resources:

• SharePoint ( .aspx pages are not supported)
• OneDrive
• Documents uploaded to Dataverse
• Custom data (internal or external): provide your own source, such as a Power Automate flow or a skill.

For more information on generative responses, see: Generative responses.

Website and SharePoint URLs

To provide your agent with a broader range of knowledge, you can provide URLs to different websites and SharePoint sites. The URL is used to represent the scope of content that will be used to generate responses. To maximize the amount of data your agent can access, several points should be considered:

• URLs can have up to two levels of depth/subpaths indicated by a slash.
  Examples of valid URLs:
• www.contoso.com
• www.fabrikam.com/engines/rotary
• www.fabrikam.com/engines/rotary/

Example of invalid URL:

• www.fabrikam.com/engines/rotary/dual-shaft

Although you are limited to two subpaths in the URL, this does not mean you are limited to two subpaths in your results. Any publicly visible content in the URL you specify (including subpaths under a top-level domain) generates content for your agent. For example, if you enter www.fabrikam.com as the URL, data from www.fabrikam.com/engines/rotary and www.fabrikam.com/engines/rotary/dual-shaft will be examined for possible return as results.

Another point to consider: how to specify the domain. If you enter www.fabrikam.com (with www), only content from www will be returned. Content located on news.fabrikam.com (without www) will not be used because news is a subdomain under the top-level domain fabrikam.com.

If, instead, you enter fabrikam.com, then content from www.fabrikam.com and news.fabrikam.com will be used because they are both under the top-level domain fabrikam.com.

Other points to consider:

• Social media and forum URLs: your agent could generate nonsensical, irrelevant, or inappropriate responses if you use a forum or social network as a URL.
• Search engine URLs: do not include search engine URLs like bing.com, as they do not provide useful answers.
• SharePoint: SharePoint URLs can be added.

It is recommended to omit https:// in the URL. Recognized SharePoint URLs will come from the sharepoint.com domain. SharePoint site URLs cannot exceed two levels of depth. Content from .aspx files on SharePoint will not be used to generate responses.

For more information on URLs, see: Generative responses.

Document upload

Another option you can use as a data source for generative responses is to upload your own documents for your agent. The documents will be used throughout your agent; however, you have the option to specify nodes that should not use the uploaded documents.

Once uploaded, when a user asks a question and the agent has no defined topic to use, the agent generates a response from your uploaded documents. The agent uses generative AI to answer the user’s question and provides a response in a conversational style. Uploaded documents are securely stored in Dataverse. The number of documents you can upload is only limited by the storage space available for your Dataverse environment, and the maximum file size per document is 512 MB.

Image, audio, video, and executable files are not supported. See: Generative responses for the full list.

Orchestrate agents with generative AI

Traditionally, virtual agents relied heavily on predefined trigger phrases and hard-coded actions to navigate conversations. This often required users to phrase their requests in a specific way to get the desired response from the agent. With advances in generative AI, agents can now dynamically determine the appropriate topics and content to communicate, leading to more natural and fluid interactions.

In the Generative AI settings of Copilot Studio, you have enhanced control over how your agent interacts with users. Under the section How should your agent interact with people?, you can choose between two interaction modes:

Classic

The traditional approach to agent creation, focused on predefined triggers, topics, and actions.

• Actions can only be invoked within a specific topic.
• Topics are triggered only when a user query matches a predefined topic trigger.

Generative (Preview)

A modern approach to conversation flows that leverages generative AI.

• The agent autonomously decides which topics, actions, and knowledge to use to respond to user queries.
• Enables more natural and contextual interactions without requiring exact trigger phrases.

Using generative AI to determine how your agent responds can make the conversation more natural and fluid for users.

When a user sends a message, your agent selects one or more actions or topics to prepare its response. Several factors determine this selection. The most important factor is the description of topics and actions. Other factors include the name of a topic or action, input or output parameters, as well as their names and descriptions. Descriptions allow your agent to be more precise when matching a user’s intent to actions and topics.

In generative mode, an agent can select multiple actions or topics at once to handle multi-intent queries. Once actions and topics are selected, the agent generates a plan that determines their execution order.

For more information on generative mode, see: Orchestrate agents with generative AI.

Content moderation

Content moderation under Generative AI in the settings menu is where you can define the relevance of generated responses. You have three options to choose from:

• High – More accurate (default): The agent generates fewer responses, but they will be more relevant.
• Medium – More balanced: The agent generates more responses, but they may be less relevant.
• Low – More creative: The agent generates the largest number of responses, but they may contain inaccuracies.

For more information on content moderation, see: Generative responses.

Use generative responses in all supported languages

When creating agents, you have the option to decide which primary language your agent should use, but what happens if the user does not speak that language?

With generative response nodes, you can handle any user query you did not anticipate, such as a case where a user speaks another language that is not configured in your topics.

Generative responses can use all languages supported by Copilot Studio, except Traditional Chinese. See the article Supported languages for the list of languages and their support levels.

To configure a generative response node for the use case of a user speaking another language, start by enabling Generative orchestration on the agent’s Overview page.

Next, go to your agent’s Topics page, select + Add a topic, and choose From scratch.

Name the topic Other language (or something similar) and edit the trigger description to include the following:

This topic handles user queries in a language other than English.

Select the + sign under the trigger node and choose Advanced > Generative responses.

In the generative responses node, select Edit to display the Create generative responses properties window.

Next, enter the following prompt in the text box under Content moderation level:

If the entered language is other than English, respond with the following line: “Sorry, but I can only continue this conversation in English” in the input language.

1. Now select the input value for the generative answers node. For this example, we’ll be using the user’s last sent message.
2. Choose LastMessage.Text under the System tab from the Select a variable menu.

Finally, add an End conversation topic management node to end the conversation after the generative responses node.

Now, when a user interacts with the agent in a language other than English, the agent will inform the user that it can only continue the conversation in English, using the language in which the user initially made the request.

Test this feature in the Test your agent panel by sending a greeting in a language other than English, for example: Hola.

Notice that the agent starts the conversation in English, but as soon as another language is introduced by the user, the Other language topic is triggered and the agent responds in the user’s native language.

With generative response nodes, the possibilities are almost endless for addressing multilingual issues. Although most languages are covered by generative response nodes, a definitive list of all supported languages can be found in the following article: Supported languages.

Next unit: Test your agents