Building natural language actions in Copilot Studio

Introduction:

Copilot Studio simplifies the process of building and extending AI copilots. It allows integration of enterprise knowledge to steer the copilot to produce business relevant response, creation of new capabilities through actions, and definition of conversation paths where you want copilot to follow a definitive script, all in a no-code environment. This flexibility empowers citizen developers to tailor AI copilots to meet specific user needs.

Defining Actions:

Actions are tasks that the copilot can perform, such as retrieving data, executing workflows, or generating insights. Prompts, which are natural language instructions, enable the creation of these actions effortlessly.

Creating Custom Actions with Prompts

Prompts allow you to define actions using natural language instructions, making it a completely no-code solution. Here are three ways to leverage prompts in Copilot Studio:

• Extend Microsoft Copilots: Build natural language prompts for Microsoft first party Copilots to perform custom tasks aligned with your business goals. For example, you can describe a specific sales report format, and the copilot will generate it accordingly.

• Build New Actions for Custom Copilots: Enhance your custom copilot's capabilities by creating prompt actions that access enterprise data in Dataverse. E.g. This grounds responses in your own data, ensuring accuracy in customer queries about products or services.

• Add a Prompt Action to a Topic: Curate scripted conversations for your custom copilot using Topics. Incorporating natural language prompts to add generative actions. For instance, access a user's service records to respond more effectively to service-related complaints, providing informed and personalized support.

What's New with Prompts?

Prompts are custom generative AI actions defined in natural language, created using the Prompt Builder interface. We are excited to introduce four new capabilities for prompts:

• RAG with Dataverse Tables: This feature allows makers to generate grounded responses for their prompt actions. By selecting a Dataverse table and its related entities, makers can specify the grounding data for custom actions, steering the model's responses with business data rather than relying solely on the copilot's general knowledge.

• Model Selection with GPT-4: Makers can now choose between GPT-3.5 and GPT-4 for building prompts. GPT-3.5 offers speed and quick responses, while GPT-4 is better suited for more complex actions.

• Creativity Control: Makers have control over the creativity of the model's responses. Using the temperature setting, they can decide whether they want the responses to be more deterministic or more creative.

• JSON Output (Coming Soon): Soon, makers will be able to define structured outputs for their actions. This feature will be particularly useful for converting unstructured data into structured formats for user-friendly visualization or downstream processing.

These new capabilities enhance the functionality and flexibility of prompts, making it easier for makers to create precise and effective custom actions. Prompts can be leveraged in Power Automate and Power Apps as well.

Utilizing prompts to create custom actions in Copilot Studio empowers users to enhance their AI copilots efficiently. By providing clear instructions and leveraging the no-code environment, businesses can achieve greater operational efficiency and effectiveness.

For more detailed instructions, visit the following:

Create prompt action in Copilot Studio: https://learn.microsoft.com/en-us/microsoft-copilot-studio/copilot-ai-plugins?tabs=m365#generate-content-or-extract-insights-with-ai-builder-prompts

Create a prompt: https://learn.microsoft.com/en-us/ai-builder/create-a-custom-prompt

Use your own data in a prompt: https://learn.microsoft.com/en-us/ai-builder/use-your-own-prompt-data

Model selection and temperature settings: https://learn.microsoft.com/en-us/ai-builder/prompt-modelsettings