Getting Started with Copilot Studio: A Practical Guide

A step-by-step practical guide for teams building their first Copilot Studio agent, covering what the platform actually does, first agent setup, topics and actions, testing, and deployment to Microsoft Teams and your website.

What Copilot Studio Is (and Is Not)

Copilot Studio is Microsoft's low-code platform for building AI-powered conversational agents. It sits within the Power Platform family and is designed to be built by business analysts and developers with moderate technical skills, not machine learning engineers.

What Copilot Studio does well:

• Intent recognition from natural language input (no training data required. It uses generative AI)
• Answering questions from your organisation's knowledge sources (SharePoint, websites, uploaded documents)
• Guided conversation flows with decision trees and forms
• Calling Power Automate flows to take action (update a record, send an email, create a ticket)
• Integrating with Microsoft Teams, websites, SharePoint, and other channels

What Copilot Studio does not do:

• Train custom machine learning models
• Replace Azure AI Foundry for enterprise-scale RAG architectures
• Handle unstructured, freeform data analysis
• Operate autonomously without defined topics and knowledge sources

Licensing and Prerequisites

Copilot Studio is licensed per message (approximately £0.004 per message) with a minimum monthly commitment, or via a flat-rate capacity pack. Each conversation between a user and an agent consumes messages.

What you need to get started:

• A Microsoft 365 tenant with Power Platform enabled
• A Copilot Studio licence or a trial (available at copilotstudio.microsoft.com)
• A Power Platform environment (not the default environment for production agents)

First Agent Setup, Step by Step

Step 1: Create a New Agent

1. Navigate to copilotstudio.microsoft.com
2. Select your environment (use a sandbox environment for your first agent, never build in default)
3. Click Create and then New Agent
4. Give the agent a name and description. The description is used by the AI orchestrator to understand when to route to this agent, be specific.
5. Set the agent's instructions: a system prompt that defines the agent's persona, scope, and tone.

Step 2: Add a Knowledge Source

Knowledge sources allow your agent to answer questions from your organisation's content, without you manually scripting every possible answer.

1. In the agent settings, click Knowledge and then Add Knowledge
2. Options include: SharePoint site, public website URL, uploaded files (PDF, Word, PowerPoint), and Dataverse tables
3. Add your primary knowledge source
4. Wait for indexing to complete (typically 2–10 minutes for SharePoint)

The agent will now answer questions by searching and summarising content from this source using generative AI.

Step 3: Create Your First Topic

Topics are the structured conversation paths for specific intents that require more than a knowledge answer.

1. Click Topics and then Add Topic and then From blank
2. Add trigger phrases, 5–10 natural language examples of what a user might say to trigger this topic
3. Add a Message node, the agent's first response
4. Add a Question node to collect information
5. Add a Call an Action node to trigger a Power Automate flow
6. Add a Message node with the confirmation

Entities, Making Your Topics Smarter

Entities allow Copilot Studio to extract specific types of information from user input without asking explicit questions.

Built-in entities recognise common types automatically: Person Name, Organisation, Date and Time, Number, Email, Phone Number, Location, URL.

Custom entities recognise domain-specific terms. Example: create a Department entity with values (IT, HR, Finance, Operations) and the agent will extract the department from user input without asking a separate question.

To use an entity in a topic: in a Question node, set the Identify field to the entity type.

Testing Your Agent

The Test Pane in Copilot Studio provides a real-time chat interface with your agent. Key testing practices:

• Test every topic with at least 10 different phrasings
• Test edge cases: misspellings, partial phrases, phrases that are close to another topic's triggers
• Test fallback behaviour, what happens when the agent doesn't understand?
• Use the Topic Tracking panel to see which topic triggered and which node the conversation is at
• Use Variables Tracking to verify that entity extraction and variable assignment are working correctly

Publishing to Microsoft Teams

1. In the agent settings, click Channels and then Microsoft Teams
2. Click Turn on Teams
3. Submit for admin approval if your tenant requires it
4. Once approved, users can find the agent in Teams by searching for its name
5. For organisation-wide deployment: use the Teams admin centre to push the agent app to specific users or groups

Teams-specific considerations:

• The agent appears as a chat contact in Teams, users message it as they would message a colleague
• Adaptive cards render natively in Teams, use them for confirmations, summaries, and structured data display
• Proactive messaging requires additional configuration via the Microsoft Graph API

Publishing to Your Website

1. Click Channels and then Custom Website
2. Copy the embed code snippet
3. Paste the snippet into your website's HTML. It creates a floating chat button
4. Customise the widget appearance in the channel settings

Common First-Agent Pitfalls

• Too many topics too soon, start with 3–5 topics covering your highest-volume use cases
• No fallback topic, configure the Fallback system topic to provide helpful guidance when the agent doesn't understand
• Ignoring the escalation path, always configure a route to a human agent for complex or sensitive issues
• Not setting agent instructions, the system prompt dramatically affects response quality and scope adherence
• Testing only in the test pane, test in the actual channel (Teams, website) before go-live; rendering and behaviour can differ