Imagine an AI that can think for itself and make decisions by itself. Rather than simply responding to events, an agentic AI can be trained on your needs. The idea of building agentic AI is to empower an automated system to perform regular tasks. The nature of agentic AI allows it to determine how to achieve the complex scenarios that the user may present.
Once given a task, the agent considers the different options based on the type of knowledge it was given. The goal is to provide the agent with general knowledge and guidance; we will cover this today.
Agentic AI functions similarly to an extremely well-organized and driven individual. It starts with a given request, such as “plan my next vacation.” It then breaks down into multiple necessary steps, like “find hotels, research flights, and create an itinerary.” For personal use, there are many different use cases. Today, we will explore the business use cases and how they can help you.
Now, imagine taking this idea of an agentic AI and putting it into your own hands. For businesses already deeply invested in Microsoft’s ecosystem, Copilot Studio Agents are a game-changer. It’s a tool that empowers you to build your very own AI assistants that can tackle specific jobs, learn from their interactions, and even make decisions on their own, all without you needing to be a coding expert.
Think of it as giving your AI a brain and the ability to act on its own, tailored precisely to what you need, and integrated with your existing Microsoft tools. These AI assistants are built using Copilot Agents and Copilot Flows, which provide the framework for creating intelligent, automated workflows.
What is a Copilot Agent?
93% of the Fortune 100 use the Microsoft environment internally. Thus, when picking the proper AI tool to be implemented, Copilot is the main option for this: it is already fully integrated with your working environment.
Copilo agents will work hand-in-hand with all your other Microsoft tools and information, so you don’t need to teach them everything from scratch. It’s all about making your business operations smoother without needing a team of tech gurus.
Through building a Copilot Agent, users will leverage their existing knowledge and resources to “educate” the agent you are building. Think about connecting your SharePoint/OneDrive documents to an agent. The data resides in your environment, and you won’t need to upload it anywhere outside of your organization.
So, how does this clever assistant work its magic?
However, the actual method may be different depending on how the agent is built. The overall idea is the same. You give it a task, big or small, like “process all new customer orders” or “generate a monthly sales report.” The Copilot Agent then takes that goal and breaks it down into smaller, manageable pieces.
When creating the Agent, users will be able to define how the AI breaks the task at hand down.
Does it take a more straightforward approach, like when the user needs to know how many PTO’s they have left or what a specific well-documented policy is? This would require a simpler Agent.
Does it take a more step-by-step approach, analyzing the requests and breaking them down into multiple steps, like when the user asks for a financial analysis to be performed on an uploaded document? This would require a more complex Agent.
The approach the agentic AI will take will mostly depend on how complex the agent is built. Because agents should be built based on the use case they would be working on, it then becomes a straightforward approach. Build a simple agent for simple tasks. Build a complex agent for complex tasks.
Small Disclaimers
You’re likely here because you’re interested in Microsoft Copilot Agents. As you may know, Microsoft products are not always straightforward. Copilot agents can be built from different interfaces, such as M365, Copilot Studio, and Teams (through M365). So don’t get confused; it is almost the same thing. I highly recommend building agents through Copilot Studio, as this offers the most complete view of all the settings available when building an agent. Therefore, I will show you how to build an agent using this interface.
Microsoft loves to make things more complicated than they truly are. You may hear the name Copilot in multiple places throughout the Microsoft environment.
Are they all the same thing? Not really. Microsoft is just calling the different AI tools it has by the same name. For example, the Copilot you see when you open Excel is NOT the same as the AI Agents we have discussed so far. Our focus is on the agentic AI built through Copilot Agents.
Let’s Build an Agent
The Initial Settings
For our example today, let’s use PDF data as our knowledge base. We want to rely on the AI to interpret and answer any questions related to all my invoices stored in OneDrive. They are all stored in a single folder to facilitate the example, but they don’t need to be.
Our goal is to be able to ask the AI questions like:
How many times have we sold X item in the last Y months/years
What is in the invoice X
What was the total invoice amount between a specified date range?
These questions could be answered through SQL if we had that data in a table/view format in Snowflake, or could be answered through a Power BI dashboard; however, the idea is that nowadays, we have a new option: feed an AI model and let it answer your questions.
The first step is to define what the Agent will do, its purpose. Once that is done. The goal is to provide it with high-level behavioral instructions. We are looking to provide it with details on what it is to expect from the users interacting with it, and then receive instructions on how to interact back.
The idea is not to define the exact instructions it needs to go through for each possible question that will be asked. This will be done later on through Topics. The first set of instructions is to define the high-level context. Think about it, if someone needs to buy groceries, you explain to them what a grocery store is prior to explaining to them what items can be found in each aisle.
Knowledge Base
Our knowledge base will be a OneDrive folder, but realistically, it could be anything you need. You can pick one source of data you have or multiple; it doesn’t matter as long as it provides the AI with the context it needs to know.
Copilot Studio is super flexible. We can give the AI agent lots of different kinds of information. You can easily connect it to a OneDrive folder, like we did above, but you can also link it to things like your company’s SharePoint sites, websites that are open to everyone, or even special data sources you’ve set up yourself, like your company’s Snowflake or other databases.
This means you can teach your AI exactly what it needs to know, whether it’s about company rules, what products you sell, or information that changes often on the internet. The main idea is to pick the best way (or ways) to give your AI the right and most useful information for what you want it to do.
For the documentation on the details for connecting to OneDrive:
Add OneDrive files and folders as a knowledge source | Microsoft Learn
We have loaded 100s of these types of files (image below) to a OneDrive folder. All of these files will be indexed through Dataverse. Keep in mind that he speed you can index these files will vary depending on the type of license you have. If you have the Copilot Pro license, you will be able to index larger files faster.
Topics
How do we add complexity to an Agent? Topics.
Topics in Copilot Studio define what an AI agent knows and how it should use that information. They allow you to teach the agent specific subjects by linking to knowledge bases (like documents), deciding if it can search the web, and guiding how it answers questions.
By default, any new agent will come with a set of Custom Topics (mostly examples for you to leverage) and System Topics (these make the engine run). Changing Custom Topics shouldn’t break your agent; changing System Topics may impact it. Users should change System Topics when they feel comfortable working with topics. You can create new topics as well.
Topics allow you to refine how the Agent should behave; thus, by building multiple topics, you can control the flow of thoughts with more precision:
Let’s say you have an Agent who knows all your HR policies. If you store the AI instructions in the same topic, you are asking the AI to use ALL its knowledge to interpret the question, go through the knowledge base, and retrieve the answers. This method leaves you susceptible to incorrect information if your knowledge is not “clean” or if the question is too ambiguous.
However, in this example, let’s say that you are able to define a Topic for PTO policies, another Topic for reimbursements, another for employee conduct, another for benefits, etc…
Remember the grocery example I used earlier? Even with the proper high-level context instructions given earlier, someone might still wander around, lost in the grocery store, looking for what they need. What if we could teach them earlier that a grocery shop has different sections? If you need fresh ground beef, you head to the meat department; if you need cleaning supplies, you go to the cleaning department.
Providing a clear path to the AI will result in faster and cleaner answers. For instance, if a user inquires about a specific holiday, the AI will not search through the employee conduct knowledge base. Instead, it will go directly to the Topic related to PTO and Holidays.
For this blog’s example, as explained above, we are looking to build an agent that retrieves any invoice information from our invoice knowledge base. For this, I don’t need multiple topics because I have a single type of knowledge base, a folder in OneDrive with all my invoices.
We will cover some necessary steps for creating a new topic:
The details tabs will allow you to define the model instructions, and by using the Input tab, you can define the user’s original request as the command to be interpreted by the agent
Model description is where you will insert the trigger phrases. This means the Agent will use those examples as the “rule” to define whether that topic is to be used or ignored.
The Trigger is what controls the usage. Make sure it contains good examples or a proper description of what should activate the trigger. Remember this is an AI interpretation of the user’s message, so it does not need to be exactly word-for-word, but the AI should be able to connect what the user is requesting to what this topic accepts.
If you need the topic to answer back using an AI model, then you set it up through a Create Generate Answer.
Below, we will dive deeper into what truly matters for the topic, which is the ability to answer your questions.
The main goal of the topic is to provide the context to the Agent and allow the agent to answer it properly. What does it mean? This means you are defining exactly what the AI should know and use to “educate it” and then allow it to answer your questions with this “education”.
-
By creating a generative answer, you will define the step of answering the question you asked.
-
You can link specific knowledge bases to it. In this case, we are linking the only knowledge base we uploaded to the agent, our folder with the invoices.
-
Should the agent be able to search the web to complement its knowledge? In this example, no. I want it to use ONLY the invoices.
-
Should the agent be able to use its own knowledge? Again, in this example, I do not. The only knowledge and understanding it needs will be given on step 2 above and step 5 below.
-
The prompt. This is how we give the AI the instructions on how to answer. I highly recommend using the known LLM models to help you write the best instructions possible.
Testing the Agent
Below, I will include the screenshot of a random invoice in the hundreds we have in the folder, as well as an example where I’m asking the agent to help me retrieve information from a specific invoice I need to look up.
PDF File:
Copilot Interaction:
The copilot agent was able to retrieve all the information found in the invoice. We will also add a reference at the bottom in case we want to open it directly. It is important to note that even though the file name contains a generic name, it was not used as the search engine’s parameter. As mentioned earlier, Copilot indexed all these files and then stored the individual information.
Does using good and clean file names help? Yes, but it is not a deal breaker if you set up the topic’s knowledge the correct way.
When testing out the agent, it is good to have the “Track between topics” option enabled. This way, you will be able to follow how the agent is thinking about your request. This way, you can make sure the topic’s trigger is correctly set up, thus the request would activate the correct topic, which would then leverage the correct knowledge base. Microsoft is changing the layout often; currently, this option is found in the ellipsis in the Test view.
Going back to the grocery example. If the user is asking about where to pick up some ribeye, I don’t want it to search for it in the soda aisle. This feature allows the developer to confirm that the correct topic is getting triggered.
Deployment
This agent can be deployed in any of the Microsoft tools that currently accept an agent, Teams, or the M365 page (recommended), or SharePoint as well. It can also be deployed externally in a website. Below are the current available options. Keep in mind that this list is growing, and in the future, more interesting options will be available.
Most companies would have the idea to deploy to Teams or the M365 platform (or both). To do so, we will deploy the agent by clicking on the option and going through the steps (currently, it requires signing in and clicking on the Add to Channel button). Once that is done, we can see it in either Teams or M365.
Under the availability options, you can control the proper access. Should the whole company have access to the agent? Should a small group of individuals? We can use Azure Active Directory options to better control access.
The magic of these agents is how they can take on those complex, repetitive tasks, tap into all the knowledge your organization has stored, and then give smart, relevant answers. Ultimately, this means smoother operations and way more efficiency within the whole Microsoft world.
So, as you start building your own Copilot Agents, just keep in mind how flexible and scalable Copilot Studio is. Whether you’re making something super simple for internal questions or a really sophisticated agent for heavy-duty data analysis, the framework lets you create exactly what you need. Plus, Microsoft’s AI tools are always getting better, promising even more powerful features and ways to deploy in the future, so make sure you keep yourself up to date with the latest updates. That makes really understanding and using Copilot Agents a fantastic investment for any organization that wants to truly leverage AI.
Conclusion
In this article, we looked at how to build a Copilot Studio Agent and how to integrate it into your existing Microsoft environment. Copilot Agents are great resources for this type of approach because they offer the ability to create agents and are fully integrated with Microsoft.
We walked through a meaningful example I have seen dozens of companies need to go through: the extraction of information stored in a local enterprise repository.
Looking for ways to make your business more efficient with AI-powered tools?
By using agentic AI in your organization, users can leverage the power of LLMs and the user’s (or company’s) own resources to achieve better results.
Our team can help you identify the highest-impact use cases, design a right-sized pilot, and prove value on your own data and workflows.
