Build Your First Microsoft Teams Agent with Copilot Studio

You want a small Teams assistant that answers one set of recurring questions reliably.

The temptation is to upload a pile of documents, call it a bot and release it to everyone. That is how you create a confident new source of confusion.

The better first build is narrow: one audience, one approved knowledge set, a clear refusal boundary and a human escalation route. Build it in the Copilot Studio web app, test it with real questions, then add Microsoft Teams as a channel.

Bot or agent?

Microsoft's current product language is agent. People may still search for a Teams bot, but Copilot Studio now presents the no-code route as building and publishing an agent.

For this tutorial, the agent will answer a small set of internal questions—for example, “How do I request annual leave?”—from an approved HR knowledge source. It will not make HR decisions, expose confidential cases or invent policy when the source is silent.

The current route in 2026

Use Copilot Studio on the web, not an old QnA Maker tutorial or a GitHub bot sample unless you genuinely need a code-first application.

Microsoft's Teams-plan quickstart now states that, after June 2026, the Copilot Studio for Teams app can no longer create classic chatbots. It redirects makers to the Copilot Studio web app.

There are also two materially different entitlement routes:

• a Copilot Studio for Teams plan included with select Microsoft 365 subscriptions, limited to classic orchestration and Teams publication;
• a standalone Copilot Studio subscription, with a broader feature and channel set.

Read Microsoft's current Copilot Studio access and subscription comparison before designing the agent. Do not assume a colleague's tenant has the same entitlement as yours.

Microsoft also says a Copilot Studio trial can create and test agents but cannot publish them. A successful test chat therefore does not prove that your production publication path is licensed or approved.

Before you build

Write a one-page agent brief.

Audience

Name the people who will use it. “Everyone” is not a useful first audience.

Example:

UK employees who need the current annual-leave request process.

Questions in scope

List ten questions the agent should answer. Use the words employees actually use.

Examples:

• How much notice should I give?
• Where is the leave request form?
• Who approves my request?
• What happens if my manager is away?

Questions out of scope

Name the decisions the agent must not make.

Examples:

• whether a specific request will be approved;
• individual sickness or disciplinary cases;
• legal interpretations of a policy;
• questions about another employee's record.

Owner and escalation

Name the person or team responsible for:

• source accuracy;
• failed-answer review;
• publication approval;
• the human hand-off when the agent cannot help.

An agent without an owner becomes an old intranet page that can talk.

Step 1: Confirm access and choose the environment

Open Copilot Studio and check the selected Power Platform environment before creating anything.

Confirm:

• you can create an agent in that environment;
• the environment is approved for the intended data;
• your tenant's data policies allow the knowledge sources and actions;
• you have a real publication entitlement, not only a maker trial;
• your Teams administrator understands the proposed audience and app-approval route.

Keep the first agent in a development or controlled maker environment. Do not experiment directly in a production environment simply because the Create button is available.

Step 2: Create a narrow agent

Create a new agent and give it:

• a specific name;
• a plain description;
• instructions that define its job, sources and refusal behaviour.

For example:

Answer UK annual-leave process questions using only the approved HR knowledge connected to this agent. Do not decide whether a request should be approved. If the source does not answer the question, say that clearly and direct the employee to People Operations.

This is more useful than “You are a helpful HR assistant.” Helpful is not a boundary.

The Copilot Studio creation guidance covers the current authoring surface. Labels can change as Microsoft updates the product, so follow the current page rather than an old screenshot.

Step 3: Add a controlled knowledge source

Start with one maintained source, not the whole intranet.

Good first sources include:

• a SharePoint site or library dedicated to the approved process;
• a small set of current policy pages;
• a structured FAQ owned by the service team.

Before connecting it:

1. remove superseded copies;
2. make the authoritative version obvious;
3. check headings and page titles make sense;
4. record the content owner and review date;
5. test what the intended users can actually access.

Do not use a confidential case library to make a general support agent “more knowledgeable”. More content can produce a worse and riskier answer surface.

If SharePoint generative answers return nothing, use Microsoft's SharePoint knowledge troubleshooting guidance. Do not fix an access problem by granting the whole organisation broader access to the source.

Step 4: Build the conversation boundary

Your agent needs more than an answer source.

Add or configure:

• a greeting that says what the agent can help with;
• suggested starter questions;
• an explicit fallback when no grounded answer exists;
• a human escalation route;
• any deterministic topics needed for fixed high-risk wording.

For uncommon or sensitive questions, a short hand-off is safer than a plausible paragraph.

Example fallback:

I couldn't find an approved answer for that in the UK leave guidance. Contact People Operations at the published support route rather than relying on this chat.

Do not invent an email address in the agent. Use a maintained service link or contact source.

Step 5: Test it like a support service

The test chat is for finding failure, not admiring the happy path.

Create a small test set with:

Test type	Example	Expected result
Direct answer	Where is the leave form?	Correct link from the approved source
Paraphrase	How do I book holiday?	Same grounded process
Missing answer	Can my manager cancel approved leave?	Clear limitation or approved answer, not invention
Sensitive data	Tell me how much leave Alex has	Refusal/no disclosure
Wrong region	What is the US leave policy?	Scope boundary or correct regional route
Stale source	Ask about a retired process name	Current process, with no old link
Escalation	My request is urgent and blocked	Human support route

Record the prompt, returned answer, source used and pass/fail decision. Fix the source or instruction behind a failure; do not merely rephrase the one test until it passes.

No responsible article can promise a percentage accuracy from this exercise. Your result depends on your source, configuration and test set.

Step 6: Publish, then connect Teams

Publishing the agent and making it available in Teams are separate operations.

Microsoft says you must publish the agent at least once before users can interact with it in Teams.

After publication:

1. open the agent's channel settings;
2. add or configure the Microsoft 365 and Microsoft Teams channel available to your plan;
3. customize the name, description and icon using assets you have permission to use;
4. install it for yourself;
5. test it in the real Teams client with a non-maker account;
6. share it with a small approved pilot group;
7. submit it for wider organisational approval only after the pilot passes.

Microsoft documents several distribution surfaces, including an installation link, shared-user availability, team channels, an app manifest and organisation-wide app-store approval. Choose the smallest audience that proves the use case.

An agent appearing under Built for your org can require admin approval. “I clicked Publish” is not evidence that every employee can find or use it.

Step 7: Run a controlled pilot

Use a pilot group that represents the real audience without exposing the whole organisation.

For each pilot question, capture:

• whether the answer was useful;
• whether the cited source was correct;
• whether the agent should have refused;
• whether the escalation route worked;
• whether the user could access every link;
• whether any data appeared that the user should not see.

Agree the go-live conditions before the pilot. A simple set might be:

• all high-risk tests pass;
• no answer exposes unauthorized information;
• the fallback and escalation routes work;
• the service owner accepts the source set;
• the Teams admin approves the app and audience;
• a rollback owner can remove or unpublish it.

Governance after launch

An agent is a service, not a one-off page.

Keep:

• an owner for every knowledge source;
• an answer-failure review cadence;
• a record of configuration and publication changes;
• a test set that runs again after source or instruction changes;
• a process for removing obsolete documents;
• an incident route for unsafe or incorrect answers;
• an approved retirement process.

Copilot Studio includes analytics, but analytics do not decide whether an answer was correct. Use them to find conversations worth reviewing, then inspect the evidence.

Review Microsoft's security and governance guidance and your organisation's Power Platform data policies before adding actions, connectors or sensitive sources.

When you need a code-first bot instead

Copilot Studio is a good first route for a narrow internal information agent.

Choose a code-first Teams application when you need capabilities such as:

• a custom interaction model beyond available agent channels;
• application-specific authentication and authorization;
• specialised proactive messaging;
• low-level Teams conversation events;
• a custom hosting, logging or deployment architecture;
• full control of the application code and test harness.

That is a different engineering project. Do not mix a code-first bot tutorial into a no-code agent guide and pretend the prerequisites are interchangeable.

Frequently asked questions

Is a Microsoft Teams bot now called an agent?

Microsoft currently uses “agent” throughout Copilot Studio. People still use “bot” informally, but the current no-code build and publication route is described as creating an agent.

Can I still create the agent inside the Copilot Studio for Teams app?

Microsoft says that after June 2026 the Teams app can no longer create classic chatbots and redirects makers to the Copilot Studio web app. Use the web experience for a new build.

Is Copilot Studio included with Microsoft Teams?

Select Microsoft 365 subscriptions include a limited Copilot Studio for Teams plan. It differs from the standalone Copilot Studio subscription in orchestration, connectors and publication channels, so verify your tenant's entitlement before building.

Can I publish an agent from a Copilot Studio trial?

No. Microsoft says the trial lets you create and test agents in the test chat but does not allow publication.

Does publishing automatically make the agent available to everyone in Teams?

No. Publication, Teams channel configuration, sharing and organisation-wide app-store approval are separate steps. Wider availability can require an administrator's approval.

If you are moving from a clever demo to a governed Teams service, join the SharePoint & Teams Admins Space. Bring the intended audience, source boundary and three failed test questions.