How to Use AI Builder Prompts in Copilot Studio (Demo)

Want to see how to use AI Builder prompts inside Copilot Studio to transform data using one of the 11,000 models from Azure AI Foundry? This tutorial shows you exactly how to do that — and it’s straight from my full Pluralsight course

Watch the full course on Pluralsight

In this video, you’ll learn how to:
Build an AI prompt that pulls vacation data from Dataverse
Choose from over 11,000 models, including GPT‑4.1, GPT‑5, and even Grok or Gemini
Dynamically inject user-specific data into your agent’s responses
Turn your Copilot Studio agent into a true assistant — not just a chatbot

Whether you’re building HR self-service agents, automating support, or exploring what’s possible with AI in Microsoft 365, this hands-on demo gives you the tools to get started right.

Video SummaryAI in Microsoft 365 is evolving — it’s not just about generating text anymore. It’s about creating intelligent agents that can take real action, like pulling live data to help users.Copilot Studio makes it easy to build smart agents — in the demo, the speaker adds a topic called “vacation days” that responds based on user intent, not just trigger phrases.Agents can connect to real databases — instead of giving generic info, the agent pulls personalized vacation day data from a Dataverse table, which could also be connected to systems like Workday or even Excel.You can choose from thousands of AI models — over 11,000 models are available, including GPT versions and others like DeepSeek or Gemini, allowing you to tailor the agent’s capabilities to your needs.The result is a truly helpful assistant — the agent dynamically retrieves vacation day info based on the user’s email and responds with accurate, personalized data, turning a simple chatbot into a powerful tool.

For more information, read the transcript blog below, or watch the video above!

Transcript

AI in Microsoft 365 isn’t just about Copilot writing text—it’s about agents that take action. In this demo for my Pluralsight course, I demonstrate how to use AI prompts in Copilot Studio to retrieve real data using your model. Let’s jump right in. I’m continuing to build our employee self-service agent, and for this next demo, I want to add a topic. I’m adding one from blank, and you’ll notice something different: the agent chooses based on intent. I don’t need a list of trigger phrases. I just explain what the topic does, and the agent will redirect the user based on that intent.

For this topic, I’m calling it “vacation days.” The description includes queries like “vacation days,” “how many vacation days do I have,” “check my vacation balance,” “PTO,” and so on. You can give it a list of phrases or just describe when to use the topic, like when the user wants to know how many vacation days they have left. It’s really up to you, and I recommend experimenting with your own scenarios because every model and situation is different.

Earlier in the course, I tested asking about vacations, and the agent gave general info from my knowledge sources—like “if you’ve worked between 0 and 3 years, you get this many days.” But that’s not good enough for my organization. I have a Dataverse table that logs how many vacation days each user has left. You might have this data in Workday or another HR system—or even in Excel (hopefully not, but it would still work). I want my agent to go into this Dataverse database, get the number of days I have left, and send me that info.

To do that, I’m using a new prompt from AI Builder. I name it “prompt vacation days.” When I go to the model, I see five default choices: GPT-4 mini, GPT-4.1, GPT-5, GPT-3, and GPT-5 Reasoning. But I can also add models from Azure AI Foundry, which has over 11,000 models to choose from. So if you want to use DeepSeek, Gemini, Grok, or others, you can. That’s amazing.

In my prompt, I tell it to search the database and add the table from Dataverse. There’s no search function, so I scroll to find my “employee holidays” table. I keep the user email and the “days left this year” columns. Then I add a text input called “user email” and use sample data—Vanessa, for example. I tell it to only return the number. When I test it, I get 35, which is correct. For this scenario, I probably don’t need a specialized model like Grok 3, which would add cost. GPT-4.1 works fine and also returns 35.

This shows how you can choose the right model for your business needs. Your agent can use 20 different models if needed, depending on their skills. I test it again with Vlad’s email, and it returns 50—perfect. I save the prompt and go back to the topic. Now I need to give it the dynamic user email. I create a new variable called “days left this year” and save it. That stores the info, but it doesn’t tell the user yet. So I add a message: “You have [days left this year] vacation days left.” I save and test it.

I type “vacation days,” and it hits the topic. It runs the prompt, and since I’m logged in as Vlad, it tells me I have 50 vacation days left. That’s awesome. Now the agent can look up info in a Dataverse table, in addition to all the knowledge it has from my intranet and benefits site. We’ve empowered it to connect to another system and help the user. And there’s even more we can do. That’s how AI prompts can turn a simple agent into a real assistant.

It’s been a wonderful journey—from building to publishing and monitoring agents. Check out the full course on Pluralsight by clicking the link on the screen or in the description.