🚨 Mission 06: Create a custom agent using natural language with AI and grounding it with your data

🕵️‍♂️ CODENAME: OPERATION AGENT FORGE

⏱️ Operation Time Window: ~75 minutes

🎥 Watch the Walkthrough

🎯 Mission Brief

Welcome back, Agent Maker. This mission puts you in the command seat of the most powerful capability in Copilot Studio - creating a custom agent from scratch using only natural language… and supercharging it with your own data.

This isn’t just another chatbot. You’re building a knowledge empowered digital coworker - one that can reason, respond, and reference real enterprise info.

Your weapon of choice? Natural language. Your mission? Design, train, and test a fully customized helpdesk agent that answers IT questions using SharePoint, uploaded files, or company URLs.

Let’s build your agent from the ground up.

🔎 Objectives

In this mission, you’ll learn:

1. Understanding what custom agents are and how they differ from pre-built templates
2. Creating agents using natural language prompts with AI
3. Grounding agents with enterprise knowledge sources including SharePoint, documents, and websites
4. Learning about generative orchestration and how agents dynamically search and respond using multiple data sources
5. Building and testing a fully functional IT helpdesk agent that can answer questions from your own data

🤔 What is a custom agent?

A custom agent is a chatbot or virtual assistant that you create and design in Copilot Studio to help users with specific tasks or questions. It's called custom because:

• You decide the purpose - help users request vacation time, check order status, provide assistance with IT related questions.
• You define the conversation - what the agent says and how it should respond.
• You ground it with your own data - connect to your enterprise data through the built-in supported knowledge resources.
• You connect it to your own systems or applications - choose from connectors, flows, REST APIs and model context protocol servers.

NOTE

Think of it this way: you are building your own digital helper that can talk to users and complete tasks for them such as answering questions, collecting information required by a process, or connecting to your enterprise data.

🤖 What can a custom agent do?

A custom agent can fulfill the following:

• Ask users for information such as names, dates, or preferences.
• Save that information to a database or table.
• Look up data based on the questions asked and answer them.
• Work autonomously without users directly interacting with the agent.
• Trigger actions either on-demand through direct user interaction or autonomously such as sending emails or creating records.

👩🏻‍💻 Why use a custom agent?

• Saves time by automating repetitive tasks.
• Gives users a friendly, guided experience.
• Tailor it to your business or project needs.

✨ Example

You build a custom agent that helps employees request vacation leave.

It asks for their name, vacation dates, and their manager’s name, then saves the information into the designated system that managed vacation requests, such as a SharePoint list.

Now, instead of navigating to the SharePoint list and creating a new item, employees simply chat with the agent instead.

🗣️ Use natural language to create agents

Previously you learnt how to quickly build agents in Copilot Studio using prebuilt agent templates in Lesson 05 - Get started quickly with pre-built agents. In this lesson, we'll dive into the conversational authoring experience with AI. In Copilot Studio, you don’t need to write code to create an agent. Copilot Studio makes it easy to build agents by starting with a description in your own words (natural language).

When you start by describing your agent in natural language, the AI automatically generates the agent’s name, description, and instructions. It also proposes triggers, channels, knowledge sources, and tools. You can accept or ignore these suggestions, but they only last for the current session and won’t be saved.

🌱 But I'm new to "describing what I want" - what do I do?

Describing in natural language to create a custom agent might be a new concept for you. Whenever you use Copilot across Microsoft products and services, you are using natural language in the form of a prompt.

A prompt is the message or instruction you give to an AI agent to tell it what you want it to do. Think of it as giving directions to an assistant. The clearer your instructions are, the easier it is for AI to understand and act on them.

🪄 Why Prompts matter

• They guide the agent’s behavior.
• They help the agent understand what kind of conversation to have.
• A good prompt makes the agent more helpful and accurate.

📝 Tips for writing a good prompt

• Be clear and specific - say exactly what you want the agent to do.
• Think like the user - what will the user say? What should the agent reply?
• Include examples - if possible, give a sample interaction.

✨ Example

Let's say the HR team needs an agent to help with vacation requests.

The prompt could be,

I want to build an agent that helps users submit a vacation request. When a user says they want to request time off, the agent should ask for their name, the start date of their vacation, the end date of their vacation, and their manager’s name. Once the user provides this information, the agent should save it to a SharePoint list called 'Vacation Requests' and post a notification in a dedicated Microsoft Teams channel.

Why this prompt works:

• Clearly states the goal - submit a vacation request
• Describes the user interaction - what the user says and what the agent should ask
• Lists the required data - name, start date, end date, manager
• Mentions where the data goes - a SharePoint list called Vacation Requests

🔮 OK, I've created my agent... how do I next ground it with knowledge?

In Copilot Studio, knowledge sources are places where your agent can find information to give better answers. When you add these sources, your agent can pull in your enterprise data from places like Power Platform, Dynamics 365, websites, and other systems or services your company uses.

These sources work together with AI to help your agent respond more accurately to user questions, this is achieved through what is known as generative orchestration.

🌿 What is generative orchestration in the context of agents?

Generative orchestration means the agent uses AI to dynamically decide how to answer a question by combining its built-in language skills with information from your added knowledge sources.

When a user asks a question, the agent:

• Understands the question using AI.
• Can ask users for missing information by generating questions on the fly.
• Selects the most relevant knowledge sources.
• Searches those sources for answers.
• Generates a natural, helpful response using the information it found.

🏦 Why knowledge sources matter?

1. Smarter answers - when you add knowledge sources, your agent can give better, more accurate answers using real data from your organization.

2. Less manual work - you don’t have to write every possible response. The agent can search through your added sources and respond automatically.

3. Use trusted information - your agent can pull answers from systems you already use such as Dataverse, SharePoint, or company websites so that users have reliable information from a source of truth.

4. Works with generative AI - knowledge sources and AI help your agent understand questions and respond naturally, even if the question wasn't pre-programmed or added as a starter prompt.

5. Flexible and expandable - you can add knowledge sources anytime during set up or at later point in time, your agent grows smarter as your needs change.

✨ Example

Imagine you build an agent to help employees with HR questions. You add your company’s HR policy document and SharePoint site as knowledge sources.

When an employee asks, “How many vacation days am I entitled to?”, the agent uses generative orchestration to search those sources and reply with the correct policy without you having to write that answer manually. This saves you time in having to account for every possible question an employee may ask regarding their entitlements.

Types of knowledge sources that can be added

1. Public websites

   • What it does: Searches specific websites (like your company’s site) using Bing.
   • Why it’s useful: Great for pulling in public-facing info like FAQs or product details.

2. Documents

   • What it does: Uses documents that you upload directly to your agent, such as PDFs or Word files. These uploaded files are stored securely in Dataverse.
   • Why it's useful: Enables your agent to answer questions based on internal guides, manuals or policies.

3. SharePoint

   • What it does: Connects to SharePoint folders or files powered by Work IQ.
   • Why it's useful: Ideal for accessing team documents, HR policies, or project files stored in SharePoint.

4. Dataverse

   • What it does: Uses structured data from your Dataverse environment tables and rows, and can apply synonyms and glossary definitions for tables and columns for improving agent responses.
   • Why it's useful: When you need to look up enterprise data stored in Dataverse such as customer information.

5. Real-time knowledge with connectors

   • What it does: Lets your agent access live data from other enterprise systems such as Salesforce, ServiceNow, Dynamics 365, AzureSQL, Databricks, and more during a conversation, using the user's own permissions.
   • Why it's useful: It provides up to date, secure, and accurate responses without storing or duplicating data, making your agent smarter and safer.

6. Azure AI Search

   • What it does: Allows your agent to search through large sets of documents stored in Azure using semantic and vector search to understand user questions.
   • Why it's useful: Delivers accurate, trustworthy answers from complex data sources, supports citations, and scales well for large document collections with secure access controls.

🔒 Note on security

Knowledge source authentication

Some sources such as SharePoint and Dataverse require user authentication. This means the agent will only reference data in its response that the user is allowed to see. Whereas other sources may have additional configuration required for the agent to connect to it such as Azure AI Search which requires an Azure account and specifying an authentication type.

Improving your agent's responses in Copilot Studio

After your agent is provisioned from the conversational authoring experience, you'll want to test your agent against the instructions AI generated from your prompt. Improving your agent’s responses in Copilot Studio is all about making sure it understands your goals clearly and has the right information to work with.

1. Refine the agent instructions - this is where you tell your agent how it should behave. Use clear, specific language.

   For example:

   ✅ “Act like a friendly customer support agent who explains things simply.”

   ❌ “Be helpful.” (Too vague)

2. Check the tone and language - make sure the agent’s tone matches your audience.

   You can set it to be:

   • Friendly and casual.
   • Professional and concise.
   • Supportive and patient.

3. Add or update knowledge sources - if your agent needs to answer questions about a topic, make sure it has access to the right information.

   • Add links to websites, documents, or FAQs.
   • Keep the content up to date.
   • Use clear, well-structured information.

4. Use Topics and Triggers - If your agent needs to handle specific tasks or conversations, you can create topics with trigger phrases. This helps guide the conversation more precisely. We'll learn more about this in the following lesson.

5. Test with real questions - try asking your agent the kinds of questions users might ask.

   If the answers aren’t great:

   • Adjust the system instructions.
   • Add more examples or knowledge.
   • Rephrase your questions to see how it responds.

6. Review and iterate - improving an agent is an ongoing process!

   After publishing:

   • Collect feedback from users.
   • Watch for common questions or confusion.
   • Keep refining the agent’s setup.

🧪 Lab 06: Create a custom agent in Copilot Studio

We're now going to learn how to create a custom agent that can chat over your data

✨ Use case

We'll use the same use case from Lesson 03 - Create a declarative agent for Microsoft 365 Copilot

As an employee

I want to get quick and accurate help from the IT helpdesk agent for issues like device problems, network troubleshooting, printer setup

So that I can stay productive and resolve technical issues without delays

Let's begin!

✨ Prerequisites

• SharePoint site

We'll be using the Contoso IT SharePoint site from Lesson 00 - Course Setup - Step 3: Create new SharePoint site.

If you have not set up the Contoso IT SharePoint site, please head back to Lesson 00 - Course Setup - Step 3: Create new SharePoint site.

• Solution

We'll be using the Contoso Helpdesk Agent solution from Lesson 04 - Creating a Solution for your agent.

If you have not set up the Contoso Agent solution, please head back to Lesson 04 - Creating a Solution for your agent.

6.1 Use natural language to create an agent with AI

AI generated instructions may differ across sessions

When you start by describing your agent in natural language, the AI generated name, description, instructions can vary in each session. This also applies to the proposed triggers, channels, knowledge sources and tools.

1. Navigate to the Home page of Copilot Studio and in the field, enter the following prompt which describes the IT help desk agent.

   You are an IT Help Desk assistant that helps employees resolve common IT issues and find available devices. Be polite, concise, and helpful. Use Microsoft Support as the primary source: https://support.microsoft.com (and Microsoft Learn troubleshooting if needed: https://learn.microsoft.com/en-us/troubleshoot/). Do not invent steps - if you can't verify official guidance, say so and offer safe diagnostics + escalation.
   
   For troubleshooting:
   1) Ask ONE focused question if details are missing (goal, symptom/error, app/device).
   2) Try quick fixes first (restart, connectivity, sign-in, service status).
   3) Provide numbered step-by-step instructions (short, actionable).
   4) If not resolved, offer 1-2 alternative branches.
   5) After 2-3 branches, recommend escalation and provide a "ticket summary" of symptoms + error + device/app + what was tried.
   
   For devices:
   1) Ask what type of device do they need
   
   Never ask for passwords/OTP. Refuse requests to bypass security.
   Include relevant Microsoft Support links and preserve URLs.

   The prompt covers:

   • Role and goal: IT Help Desk assistant
   • Primary knowledge sources including a hierarchy for the website knowledge sources
   • Response style: polite, concise, helpful
   • Troubleshooting flow: Question > quick fixes > steps > branches > escalation
   • Escalation artifact: ticket summary
   • Device support (basic)
   • Security boundaries: no passwords, no bypassing security
   • Link handling: preserve URLs, cite Microsoft Support

   

2. We'll double check the solution that our agent will be created in, is the solution we created and selected as the preferred solution in Lesson 04 - Create a new solution.

   Select the wheel cog icon and the Agent Settings modal will appear and we can see our solution created from earlier is selected by default. This is due to selecting our solution as the preferred solution in Lesson 04 - Create a new solution.

   

3. Select Cancel. Submit the prompt description and Copilot Studio will begin provisioning our agent.

4. Once the agent has been provisioned, you'll see a confirmation appear. Notice how AI automatically generated the name, description and instructions for your agent. The orchestration mode is enabled by default (found in Settings) and the default model is used for the response model of the agent.

   Reminder: AI generated instructions may differ across sessions

   When you start by describing your agent in natural language, the AI generated name, description, instructions can vary in each session. This also applies to the proposed triggers, channels, knowledge sources and tools.

   

5. Scroll down to review the AI suggestions for knowledge sources, tools and triggers.

   

6. Scroll down some more to review the Connected Agents, Topics and Suggested Prompts sections.

   

7. We'll next double check our agent has correctly been created in the Contoso Helpdesk Agent solution. Select Settings on the upper right.

   

8. We can see under Advanced that the agent has been created in the Contoso Helpdesk Agent solution. Nice! Exit from settings.

   

9. Now let's update the name of our agent. Select Edit in the Details section.

   

10. Enter the following as the name of the agent and Save the updated details.

    Contoso Helpdesk Agent

    

11. We'll now add the suggested knowledge sources. In the Knowledge section, select + Add for the website URL of https://support.microsoft.com

    

12. The Add public websites modal appears with the website URL. Select Add.

    

13. Add another website using the below URL and select Add to agent.

    https://learn.microsoft.com/troubleshoot/

    

14. The two website URLs have now been added as knowledge sources for our agent. Select X Dismiss to remove the second suggestion by AI.

    

15. By default the Web Search setting is enabled. Select the toggle to disable the Web Search feature as we only want the agent to use the knowledge sources we define.

    

16. Let's now test our newly created agent. In the Testing pane on the right hand side, select the new test session icon.

    

17. Enter the following question in the Testing pane.

    How can I check the warranty status of my Surface?

    

18. The Activity map will then load which shows us in real-time what path the agent is processing. In this scenario, our agent has understood the question and searches the knowledge sources using the two website URLs.

    Our agent responds with answers that are outlined as numbered step-by-step instructions, as defined in the instructions. The response has references to the https://support.microsoft.com website that the agent formed its response from. This enables users to verify the source of the answer.

    

Congratulations! You've built your first custom agent by starting with a description in Copilot Studio 🙌🏻

6.2 Add an internal knowledge source using a SharePoint site

Previously, we added a public website as an external knowledge source for our agent during the conversational creation experience. We're now going to add an internal knowledge source using a SharePoint site. This will be the SharePoint site you created during Lesson 00 - Course Setup.

1. In the Knowledge section, select + Add knowledge and select SharePoint.

   

2. Paste in the address of the SharePoint site created in Lesson 00 - Course Setup in the SharePoint URL field and select Add.

   Update the name of the SharePoint site to Contoso IT and select Add to agent.

   

3. The SharePoint site has now been added as a knowledge source with a status of Ready. The Status column will show whether the knowledge source has been loaded/connected to successfully, or if there is an issue.

   

6.3 Add an internal knowledge source by uploading a document

We'll now add another internal knowledge source by uploading a document directly to our agent.

1. In the Knowledge section, select + Add knowledge and select Upload file or select to browse.

   

2. Download this sample file and select it in your File Explorer. Select Open.

   

3. The file has been selected for upload. Select Add to agent next.

   

4. The document will be in the process of being added to the agent. Wait until the upload has completed, do not close the browser window.

   

5. The status of the document will initially show as In progress, wait until the status has been updated to Ready before testing your agent.

   

Let's now test our agent!

6.4 Test agent

We'll test our four knowledge sources by asking questions to our Contoso Helpdesk Agent.

1. Select the new test session icon in the test pane.

   Enter the following question to test our public website (external) knowledge source.

   How can I find the serial number on my Surface device?

   

2. You'll next see the agent reviewing the knowledge sources and providing a response using the website knowledge source.

   A response will be returned and notice how there are references to the web page it formed its answer from.

   

3. If you scroll down the knowledge modal in the activity map, you'll see the other knowledge sources the agent searched, which is the other website URL, SharePoint site and the uploaded file.

   However these were not used as in the Referenced sources section, the first website knowledge source was only referenced. The answer was grounded using the first website knowledge source. If you select the references, you'll be directed to the web page.

   

4. Let's now test both our SharePoint site knowledge source and document knowledge source in a single message. Enter the following question.

   How can I access our company's Contoso VPN from my device? How do guests connect to the Contoso Guest wifi?

   

5. Once again you'll see the agent reviewing all four knowledge sources to generate a response to the two questions submitted in a single message. The agent responds to both questions in a single message, and provides separate references of where it generated its response from.

   In the knowledge modal in the activity map, you'll see the SharePoint site referenced for the Question 1 in regard to accessing the Contoso VPN. You have full visibility of what knowledge sources were used to answer both questions in the activity modal.

   

6. Scroll down to the response for Question 2 in regard to the Contoso Guest wifi. Again, we'll see a response that's grounded using the uploaded file that contains the details.

   

7. In the activity modal, notice how the second website URL was also referred to but not used as one of the referenced sources.

   

8. It's always good to verify the generated response is correct. Select the document reference and a modal will appear with the text from the document that reflects the answer.

   

The agent can answer multiple questions in a single message, and search the knowledge sources + reference the knowledge sources in its response. Make sure to always verify the response is correct by reviewing the references.

✅ Mission Complete

Congratulations! 👏🏻 You've learnt how to create your own custom agent by starting with a description. Your custom agent can chat over your data from four different knowledge sources 🙌🏻

This is the end of Lab 06 - Create an agent with AI, select the link below to move to the next lesson. Your custom agent created in this lab will be used in the next lesson's lab.

⏭️ Move to Add a new Topic with trigger lesson

Welcome to the elite. You now know how to forge digital agents that speak your language, reference your data, and support your team. Keep going—your mission’s just getting started.

📚 Tactical Resources

🔗 Quickstart: Create and deploy an agent

🔗 Create and delete agents

🔗 Key concepts - Authoring agents

📺 Create a custom agent using natural language

📺 Add knowledge to your agents