Copilot Studio: Unlocking Conversational Magic with Plugin Actions

Conversational AI is transforming how we work with technology across every industry.? According to Gartner, by 2025, 80% of enterprise applications will embed a virtual conversational assistant to fulfill content fetching and content generation tasks (Gartner: The Impact of Generative AI on the Conversation AI Market, 2023).

Organizations that are adopting conversational AI are doing so for many reasons, such as enriching employee experiences, optimizing business processes, improving efficiency, and reducing development costs.

To realize these benefits, companies are rapidly adopting Microsoft Copilot Studio as their end-to-end conversational AI platform for building and extending copilots with generative AI and large language models.? Copilot Studio is at the forefront of this new era of conversational AI, allowing copilots to cover more topics than ever, while drastically reducing build time.

Development of these conversational experiences is shifting as well: from manually authored topics that miss opportunities outside of the topic scope to strategically authored topics plus generative AI to handle everything else.? Ultimately, this reduces the number of topics that developers need to build/manage and leverages AI to answer and act on behalf of the user.

While we’ve previously covered the Generative Answers feature in Copilot Studio, there is another capability that is accelerating and streamlining copilot development: Generative Actions via Plugins.?

Just as generative answers uses AI to answer user questions based on specified content, generative actions uses plugins + AI to take action for a user based on their inputs.? With this feature, we reduce the manual build-out of conversation paths and instead call the plugin to generate the conversation during runtime, gathering all the required inputs to run the plugin.

There are a few core plugin action types, including connector actions, Power Automate cloud flows, and Bot Framework skills.? Let’s see a plugin in action!

In my “People Finder” copilot, I’m going to add a new action and use the “Get user profile” connector action.

Then I name the plugin and provide a description.? I’m also turning on dynamic chaining so that the copilot can use AI to figure out when to call this plugin action during a conversation.

For inputs, this connector action only requires one: a user principal name/email.? I can either hardcode this or let the copilot fill this input.? In this case, I’ll let the copilot fill it and I’m going to tell the copilot to identify this input as type “Email”.

Lastly, I can specify how the copilot responds in the “Outputs” section or I can let AI dynamically generate a message, which is what I’ll use here.

I can now test my copilot.

I’m going to ask the copilot to tell me more about the person associated to this email.? The copilot uses dynamic chaining to call this plugin action and knows it needs the “User (UPN)” input to run, which I already provided in my original ask to the copilot.

Using the email from my previous message, the plugin action calls the Office 365 Users connector to get details on this person.

I can also ask my copilot different questions, like what this person does at our organization, and the copilot responds with information about Amber’s role and expertise.

Right after the copilot’s response above, I can ask follow-up questions.? Notice that I asked, “What is her office location?” and I did not provide Amber’s email for the input.? This is because the copilot understands the context of the ask: I first asked about Amber’s job, so the copilot knows that “What is her office location?” is in reference to Amber’s office location and so it uses the same email input to get her location.

Similarly, I can create an additional plugin action for getting the current weather.? This action can also be dynamically called based on what the user asks the copilot.

Now, if I ask my copilot to tell me about Amber, it provides details like job title, department, contact information, and location.? I can follow-up that response by asking “What is the weather by Amber?” and because it contextually knows that Amber’s location is Milwaukee and it understands that I'm asking about weather, it is able to feed "Milwaukee, WI" an input to the weather plugin action.

This results in the copilot informing me about the current weather in Milwaukee, even though I never had typed Milwaukee in as an input.

This is a simple example of how plugin actions can call different services to get information.? These plugins can also perform a task on behalf of the user, like creating a new ServiceNow ticket or adding a user to a new Microsoft Teams group.

Once we build these plugins, we can reuse them across other copilots.? They act as micro-apps that can be dynamically called by one or many copilots.

Plugins are already changing the world of copilot development by offering AI-powered conversational experiences for end users, all while reducing manual topic development and developer effort.

Try plugin actions in Microsoft Copilot Studio today!