Section 5: Administering Copilot Studio & Agents
Chapter 46: What is Copilot Studio?
Microsoft Copilot Studio is a low-code graphical platform used to create, test, and publish
Custom Agents. While Microsoft 365 Copilot is a ready-to-use personal assistant, Copilot
Studio allows organizations to build "specialists" that are tailored to specific business processes,
such as HR onboarding, IT support, or supply chain tracking.
Chapter 47: Understanding Agents (The Core Concept)
An Agent is a specialized AI entity designed to perform tasks or answer questions within a
specific domain. Think of an agent as a "mini-app" that uses natural language as its user
interface. It combines:
Instructions: Defining its persona and boundaries.
Knowledge: The data sources it can reference.
Actions: The tasks it can perform (like booking a flight or updating a CRM).
Chapter 48: Prebuilt vs. Custom Agents
Prebuilt Agents: These are ready-made by Microsoft or partners (e.g., the "Researcher"
or "Analyst" agents). They are designed for common tasks and require minimal setup.
Custom Agents: These are built from scratch in Copilot Studio. You define exactly what
data they access and what workflows they trigger, offering full control over the AI's
behavior.
Chapter 49: Agent Architecture
The architecture of a Copilot Studio agent relies on Topics and Nodes:
Triggers: The specific phrases or events that start a conversation.
Topics: The individual "conversational paths" (e.g., a "Refund Policy" topic).
Orchestrator: The AI logic that determines which topic or knowledge source to use
based on the user's intent.
Chapter 50: Connecting Agents to Data Sources
To make an agent "smart," you must ground it in data. Key sources include:
Public Websites: The agent scrapes a URL to answer questions.
SharePoint/OneDrive: The agent accesses internal files, respecting the user's existing
permissions.
Dataverse: The underlying database for Power Platform, ideal for structured business
data.
Custom Connectors: APIs that link the agent to external systems like SAP, Salesforce,
or Zendesk.
Chapter 51: The Agent Lifecycle: Development to Production
A disciplined lifecycle ensures quality and security:
1. Draft: Building the agent in a sandbox environment.
2. Test: Using the built-in "Test Chat" to refine instructions.
3. Publish: Making the agent's current version available to a channel.
4. Monitor: Reviewing analytics to see how the agent performs with real users.
Chapter 52: Publishing Agents to Teams
Once built, an agent must be "deployed" to a channel.
Teams Integration: Publishing to Teams allows users to find the agent in the Teams
App Store.
App Manifest: Copilot Studio generates a .zip file containing the icons and descriptions
required for Teams to recognize the agent.
Side-loading: Admins can manually upload the agent for specific departments before a
wide rollout.
Chapter 53: Managing Agent Permissions
Permissions follow a two-tier model:
Maker Permissions: Who is allowed to edit and build the agent (managed in the Power
Platform Admin Center).
User Permissions: Who is allowed to chat with the agent (managed via Security Groups
or individual sharing).
Chapter 54: Who Can Use an Agent?
You can control the visibility of an agent:
Individual/Group: Only specified users can see the agent.
Organization-wide: The agent is available to everyone in the tenant (requires Admin
Approval).
Unauthenticated: For external-facing bots (e.g., on a public website), users don't need to
sign in.
Chapter 55: Using Generative Answers
Generative Answers allow an agent to answer questions even if a specific "Topic" hasn't been
created.
The "Boost" Node: When the agent doesn't find a matching topic, it searches its
knowledge sources (like SharePoint) and generates a real-time summary.
Content Moderation: Admins can set the "strictness" level to prevent the AI from
giving unverified or harmful answers.
Chapter 56: Agent Analytics: Tracking Performance
Found in the Analytics tab of Copilot Studio:
Resolution Rate: Percentage of conversations where the agent successfully answered the
user.
Escalation Rate: How often the agent had to "Hand off" to a human.
CSAT: Customer Satisfaction scores collected through end-of-chat surveys.
Chapter 57: Monitoring Agent Conversations
Admins can review Conversation Transcripts to understand where the AI is failing.
Privacy Note: Transcripts are stored in Dataverse. Access to these logs should be
restricted to ensure user privacy and compliance with Purview settings.
Chapter 58: Troubleshooting Failed Agent Responses
Common failures include:
No Match Found: The user asked a question outside the agent's knowledge. Fix: Add a
new Knowledge Source or Topic.
Authentication Error: The agent doesn't have the right token to access a file. Fix: Check
the Entra ID app registration.
Latency: The data source (e.g., a slow API) took too long to respond.
Chapter 59: Testing Agents in the Studio
The Test Pane is your primary debugging tool.
Tracking: You can see which specific "Node" is currently active as the conversation
flows.
Variables: Monitor the values of variables (like UserEmail or OrderNumber) in real-
time to ensure data is being passed correctly.
Chapter 60: Best Practices for Agent Security
1. Least Privilege: Ensure the agent's connection only has "Read" access unless "Write" is
strictly necessary.
2. Review Shared Access: Regularly audit who has "User" permissions to avoid "Identity
Creep."
3. Data Policy (DLP): Use Power Platform DLP policies to prevent agents from sending
sensitive data to unauthorized connectors (e.g., blocking an agent from sending
SharePoint data to a public Gmail account).