Skip to content

AI Office Hours Recording

AI Agent Examples: Watch Us Build a Copilot Agent Live

 

Quick answer. A Copilot agent is an AI assistant, built in Microsoft Copilot Studio, that answers questions or takes actions using your own business data. In this recording, Eugene Matisko, a Power Platform practitioner at Endsight, builds one live: a CRM assistant that looks up accounts and contacts, cites its sources, and runs in Microsoft Teams. Watch it if you're deciding where to start with AI agents.

What this recording covers

Most business leaders have heard that Copilot can do more than answer prompts. The harder part is picturing what building one looks like, and where your own data fits in.

Eugene told the audience he would build the agent from scratch. The advance work was sample CRM data and a few prepared scripts, about eight hours of prep in all, and the build itself ran live. Publishing stalled partway through, and he worked through it on screen. If you want to know what building an agent is like for someone who does it every day, this recording shows the whole path, hiccups included.

Three AI agent examples from this session

The audience voted on what Eugene would build. Three ideas were on the ballot, and the CRM assistant won by about two votes.

  • CRM assistant (retrieval agent). Answers questions such as the phone number and website for a customer, or every contact at an account in a table, with links back to the source records. This is the one Eugene built.
  • HR policy assistant (retrieval agent). Answers questions like how to opt into the 401(k) or how temporary leave works, using your policy documents instead of an email to a coworker.
  • Email responder (autonomous agent). Watches for incoming email and replies to commonly asked questions on its own, within the permissions you give it.

Endsight also runs agents of its own, including one for support tickets and one that holds our developer best practices.

Inside the demo

  • 4:30 Eugene starts a live build in Microsoft Copilot Studio, without writing code
  • 8:56 How the audience vote chose the CRM assistant
  • 9:52 Three types of agents: retrieval, task, and autonomous
  • 18:30 Knowledge sources: SharePoint, Dynamics 365, Salesforce, ServiceNow, Dataverse, and more
  • 20:23 A tip for uploaded documents: write a description so the agent knows when to read them
  • 28:49 Settings that keep answers tied to your data, including turning off web search
  • 31:11 Testing before you publish: answers with citations and links back to the CRM
  • 34:34 Why an agent should answer with each person's own permissions
  • 40:25 Publishing the agent to Microsoft Teams
  • 48:10 Where to start: a user group, a ready environment, and regular meetings
  • 54:17 Q&A on Copilot Studio licensing and cost

Do you have questions?

Send us your question and we'll route it to Michelle, Eugene, or whoever on the team is best placed to answer it. If you'd like more sessions like this one, register for the next AI Office Hours. 

FAQ

What is a Copilot agent?
A Copilot agent is an AI assistant you build in Microsoft Copilot Studio. You give it instructions, connect it to your own data such as SharePoint files, a CRM, or a database, and people ask it questions in plain language. Depending on how you set it up, it can look up information, take actions in other systems, or run on its own. Microsoft describes Copilot Studio as a low-code platform for building agents.
What are some examples of AI agents a small business can build?
Three examples came up in our AI Office Hours session. A CRM assistant answers questions about accounts, contacts, and deals and links back to the source records. An HR policy assistant answers questions such as how to opt into the 401(k) or how temporary leave works, using your policy documents. An email responder replies to commonly asked questions on its own. Endsight also runs internal agents for support tickets and developer best practices.
What is the difference between retrieval, task, and autonomous agents?
Eugene Matisko works with three types. A retrieval agent answers questions from the knowledge sources you connect. A task agent does the same and can also take actions, such as creating a follow-up activity in your CRM when you ask for one in chat. An autonomous agent runs on its own, listening for a trigger such as an incoming email and acting within the permissions you give it. Retrieval is the starting point, and autonomous agents are the more advanced end.
Do I need to know how to code to build a Copilot agent?
No. Eugene built the CRM assistant in the recording without writing a line of code. Copilot Studio asked him plain-language questions about what the agent should do, then drafted its name, description, and instructions, which he refined. Technical skill matters more when you connect a system that has no ready-made connector, because that takes a custom connector or an API.
What systems can a Copilot agent connect to?
In the recording, Eugene lists public websites, SharePoint folders and lists, uploaded documents, Dataverse, Dynamics 365, Salesforce, ServiceNow, Azure SQL, and Zendesk. If a system offers a REST API, a custom connector can reach it. Microsoft documents the supported knowledge sources in its Copilot Studio documentation.
What does it cost to build a Copilot agent?
In the recording, Eugene explained that a paid Microsoft 365 Copilot license includes access to Copilot Studio, and that pay-as-you-go billing also works. Pay-as-you-go is also how you give people access to an agent without giving each of them a full Copilot license. The recording is from September 2025 and Microsoft changes its licensing, so confirm current pricing with Microsoft or your Endsight account manager.
Why do AI agents sometimes make up answers, and how do you stop it?
An agent invents details when it cannot find the answer in its sources and fills the gap. In the recording, Eugene turned off web search for the CRM assistant, told it in its instructions to say so when an answer is not in its knowledge, and checked each answer's citation before publishing. If an agent still invents details, he suggests removing and re-adding the source, and confirming the agent has permission to read the folder or system behind it.
How secure is a Copilot agent?
Security depends on how the agent is set up. Eugene recommends that an agent answer with each user's own permissions, so someone who cannot open a file or record in the source system does not see it in an answer either. Before you connect agents to systems outside Microsoft 365, make sure your environment is ready, because an agent can only be as careful as the access you give it.
How long does it take to build a working Copilot agent?
The live build in the recording ran about 40 minutes, after about eight hours of preparation, most of it spent creating sample CRM data. Your preparation depends on your data. Uploaded documents can take four to sixteen hours or more to index, and a large SharePoint library can take longer.
Can a Copilot agent run inside Microsoft Teams?
Yes. Eugene published the CRM assistant to Microsoft Teams, where it appeared in the sidebar for the group to pin and use. The same agent can sit on a SharePoint page or a demo website, and advanced setups can reach other channels.
Where should we start with AI agents?
Endsight recommends starting with a user group of curious people from across the organization, so ideas come up from everyday work. Make sure your environment is ready to connect agents to systems safely, especially when people have different permission levels. Then have the group meet regularly, decide what to build, and iterate. If you want help with any step, Endsight can set up a short consultation.
What is the best way to bring AI into everyday work?
Pick one or two tasks you do regularly, such as sorting an overloaded inbox, and hand them to AI the way you would hand them to a capable junior colleague. Describe what you want, check what comes back, and correct it. You learn how to iterate by doing it. You can also ask the AI what it could help with on your calendar today.

Speakers

Michelle Brezenski

Michelle Brezenski

Manager, Development @ Endsight

With over 20 years of IT experience and 16 years at Endsight, Michelle is a seasoned professional who has worked in different capacities within the company. Since 2014-15, she has led Endsight's development department's efforts and is passionate about data-driven decision making and leveraging technology to solve business problems. Michelle understands the importance of fine-tuning technology to help businesses thrive. Michelle holds multiple certificates from Microsoft, most of which focus on database management and analytics.
Eugene Matisko

Eugene Matisko

Power Platform Practitioner @ Endsight

With nearly 30 years of IT experience, Eugene works in Endsight’s Power Platform practice, specializing in Power Apps, multi-tier approval automation workflows, AI-driven solutions and infrastructure automation. He conducts consultations with Endsight’s clients to pinpoint business pain points, design tailored solutions, capture detailed requirements, and deliver precise cost estimations. Eugene holds multiple Microsoft certifications, including the Power Platform Solution Architect Expert (PL-600).




Jason Clause

Jason Clause

Director of Marketing @ Endsight

Jason has spent the last 18 years working in the managed IT services industry. He has been presenting cybersecurity events for the past several years. He has a bachelor’s degree in Public Relations from Kent State University and is a Microsoft Certified Professional. Jason is certified by Sophos, a worldwide leader in next-generation cybersecurity. 





About Endsight

Endsight is an IT consulting firm headquartered in Napa, California, serving organizations with 10 to 150 employees across California and Hawaii. Its AI practice helps businesses deploy Claude, ChatGPT, Gemini, and Copilot safely, with training, automation, and AI agents. Endsight also provides Microsoft 365 implementation, cybersecurity, compliance readiness, and strategic IT advising, with a help desk response time of 5 minutes or less. Endsight is a member of CompTIA.

Sources: Microsoft Learn, What is Microsoft Copilot Studio? and Knowledge sources in Copilot Studio.

Recorded September 11, 2025. Page last updated October 20, 2026.