Understanding the Copilot Studio UI
Explores the agent concept and Copilot Studio's main menus and core components (Topics, Generative Answers, Actions, Knowledge) from a practical perspective.
1. What is an agent?
In Copilot Studio, an “agent” is an AI-powered bot that converses with users and performs specific tasks. The unit formerly called a “Copilot” or “bot” is now standardized as an “agent.”
Beyond simply answering questions, an agent can autonomously decide the best action based on instructions and context. It combines language models, instructions, Knowledge, Topics, tools, and triggers to achieve goals.
📖 Reference: What is an agent?
1.1 Core properties of an agent
| Property | Description | Where to configure |
|---|---|---|
| Name / icon | The agent’s identifying information. Exposed to users. | Overview tab |
| Description | Describes the agent’s role and purpose. | Overview tab |
| Instructions | Defines the agent’s behavior in natural language. The LLM uses these instructions when responding. | Overview tab |
| Knowledge | Connects data sources the agent can reference. | Knowledge tab |
| Topics | Defines conversation flows for specific intents. | Topics tab |
| Actions | Connects external system calls, automation flows, and similar capabilities. | Actions tab |
| AI model | Selects the default LLM model the agent will use. | Settings > Generative AI |
📌 Agent = conversation + knowledge + action
An agent is not just a chatbot. It is an integrated unit that can be given a personality through Instructions, connected to knowledge through Knowledge, and perform real work through Actions.
An agent is not just a chatbot. It is an integrated unit that can be given a personality through Instructions, connected to knowledge through Knowledge, and perform real work through Actions.
1.2 Two operating modes of agents
| Mode | Description | Input/output |
|---|---|---|
| Conversational | Answers user questions in real time and provides guidance | Input: text → output: response, Adaptive Card |
| Autonomous | Automatically responds to events such as new records or form submissions to perform work | Input: trigger → output: automated action execution |
2. Main menus and components
2.1 Left navigation
| Menu | Description |
|---|---|
| Agents | View and manage the list of created agents |
| Environment selection | Switch Power Platform environments at the top (Dev / Test / Prod) |
2.2 Agent editing screen (six tabs)
When you select an agent, the tabs below are provided. Understanding the role of each tab clearly makes agent design much easier.
| Tab | Core role | Practical point |
|---|---|---|
| Overview | Define the agent’s identity — name, description, Instructions, language, AI model selection | Becomes the interpretation criteria for Topics, Knowledge, and Actions |
| Knowledge | Connect data sources — SharePoint, OneDrive, websites, files, Dataverse | RAG foundation for Generative Answers. The “source of truth” |
| Topics | Design conversation flows — trigger → nodes (questions, conditions, actions, messages) | Use for “areas that must not be wrong” |
| Actions | External integration — Power Automate, connectors, HTTP, MCP, AI Prompt | Core capability for “doing work beyond answering” |
| Publish | Publish and deploy channels — Teams, web, M365 Copilot, etc. | Responses differ before/after publishing, so always test after publishing |
| Analytics | Analyze usage — session count, resolution rate, unanswered questions, CSAT | The tab where real value appears in the operations stage |
📖 Reference: Create and edit topics · Knowledge sources · Use actions
⚠️ Most common mistake in PoCs
"I edited it, so why didn't it change?" — If you do not publish, changes are not reflected in real channels. Edits can be checked only in the test panel, while the previously published version remains in effect.
"I edited it, so why didn't it change?" — If you do not publish, changes are not reflected in real channels. Edits can be checked only in the test panel, while the previously published version remains in effect.
3. Orchestration modes: Classic vs Generative
The setting that determines the agent’s core behavior is in Settings > Generative AI. This choice controls the agent’s overall behavior.
| Category | Classic mode | Generative mode |
|---|---|---|
| Intent recognition | NLU matching based on trigger phrases | LLM understands intent from natural language |
| Multiple intents | One utterance → one topic | One utterance → multiple intents processed simultaneously |
| Topic selection | Trigger phrase similarity | LLM judgment based on the topic’s name and description |
| Action calls | Explicit node placement inside topics | LLM autonomously selects appropriate actions |
| Control | ✅ High (predictable) | ⚠️ Flexible, but requires testing |
| Suitable scenarios | Structured conversations, policy/process guidance | Flexible conversations, multiple intents, FAQs |
📖 Reference: Orchestrate agent behavior with generative AI
💡 Practical recommendation — hybrid model
The architecture that works best in practice is a hybrid approach: turn on Generative mode by default, but design core work as topics.
The architecture that works best in practice is a hybrid approach: turn on Generative mode by default, but design core work as topics.
- First: attempt topic matching → if matched, run the structured flow
- Second: if matching fails → Generative Answers (Knowledge-based AI answer)
- Third: if that still fails → Fallback Topic (escalation guidance)
4. Topics in detail
Topics are the core unit that defines conversation flows for specific user intents.
4.1 Three ways to create topics
| Method | Description | Best suited for |
|---|---|---|
| Manual creation | Add nodes one by one from an empty topic | When fine-grained control is needed |
| Natural-language creation (Generative Builder) | AI creates it when you describe “create a topic like this” | Fast prototyping |
| Modify an existing topic | Modify an existing topic in natural language with Copilot assistance | Iterative improvement |
📖 Reference: Create and edit topics with Copilot
4.2 Components of a topic
- Trigger: The topic is activated when the user enters specific phrases.
- Classic mode: matches by trigger phrase similarity
- Generative mode: the LLM automatically matches based on the topic’s name and description
- Node: A processing step inside a topic
| Node type | Description | Example use |
|---|---|---|
| Message node | Shows text/cards to the user | Guidance messages, result output |
| Question node | Requests information from the user (choices, text, etc.) | Name, type, date input |
| Condition node | Handles conditional branching (if/else) | Different processing by type |
| Action node | Calls external systems | API calls, email sending |
| Topic redirect | Switches the flow to another topic | Modular conversation structure |
| Generative Answers node | Generates Knowledge-based AI answers | Dynamic responses from specific sources |
4.3 System topics vs user topics
| Category | Description | Examples |
|---|---|---|
| System topics | Provided by default. Can be modified but not deleted. | Greeting, Escalation, End of Conversation, Conversational Boosting, Fallback |
| User topics | Custom conversation flows created directly by users. | Leave requests, IT requests, product inquiries |
4.4 Conversational Boosting — the built-in “AI answer engine”
When you create an agent, a Conversational Boosting system topic is automatically created. This topic includes a Generative Answers node by default, so AI answers start as soon as you connect Knowledge sources.
User question → attempt topic matching
├─ Matched → run that topic (structured flow)
└─ Not matched → Conversational Boosting Topic
→ Generative Answers node (Knowledge-based answer)
├─ Answer generated successfully → respond with citation
└─ Generation failed → Fallback Topic (escalation/guidance)
💡 Practical tip — topic design principles
- You do not need to create a topic for every question. Create topics only for structured business processing (leave requests, IT requests, etc.), and leave the rest to Generative Answers.
- In Generative mode, the topic name and description are more important than trigger phrases. The LLM uses them to decide which topic to call.
- "Topics if it must not be wrong, Generative if it must be flexible" — establish this decision criterion first.
5. Generative Answers in detail
Generative Answers is a feature that provides LLM-generated answers based on connected Knowledge sources for user questions that do not match predefined topics.
5.1 How it works
- User enters a question
- Attempt topic matching → matching fails
- Search for related documents/information in Knowledge sources (semantic search)
- Pass search results to the LLM (RAG pattern)
- LLM generates a natural answer that fits the context
- Validate results and check content safety
- Provide the answer with a citation
📖 Reference: Use generative answers in a topic
5.2 Where to place Generative Answers
Generative Answers nodes can be placed in multiple locations, allowing flexible answer strategies.
| Placement | Description | Use scenario |
|---|---|---|
| Conversational Boosting Topic (default) | Runs automatically when topic matching fails | General FAQs, document-based Q&A |
| Inside a specific topic | Inserted in the middle of a topic flow | Provide additional information in a specific context |
| Before Fallback | Last attempt before escalation | Minimize unresolved questions |
5.3 Key settings
| Setting | Description | Reference |
|---|---|---|
| Limit response scope | “Search only selected sources” — use only specific Knowledge sources | Learn |
| Use general knowledge | Use general knowledge — whether to use general information outside Knowledge | Learn |
| Content moderation | Set the strictness level of responses (high/medium/low) | Learn |
6. Actions in detail
Actions are capabilities for an agent to interact with external systems. This is the point where a “Copilot that answers” becomes an “agent that handles work”, and in most customer PoCs it is the highest-impact capability.
6.1 Action types
| Type | Description | Example | Reference |
|---|---|---|---|
| Power Automate flow | Calls an existing PA flow | Approval requests, email sending | Learn |
| Connector action | Directly calls standard/custom connectors | SharePoint lookup, Outlook mail | Learn |
| HTTP request | Direct REST API call | External ERP/CRM data | Learn |
| MCP server | Connects a Model Context Protocol server | Calls tools from a custom MCP server | |
| AI Prompt | Runs a custom prompt against an AI model | Generate HTML cards, summarize data | Learn |
6.2 Action execution in Classic vs Generative mode
| Category | Classic mode | Generative mode |
|---|---|---|
| Call method | Explicit call from an action node inside a topic | LLM autonomously decides and calls |
| Parameter collection | Collected in advance with question nodes | Asks the user only for missing information |
| Result handling | Stored in variables and used manually | Automatically summarized and reflected in the response |
📖 Reference: Orchestrate agent behavior with generative AI
⚠️ Caution when using Generative Actions
In Generative mode, the agent may call actions unintentionally. For actions with side effects such as sending email or modifying data, clearly restrict the call conditions in the instructions. Example: "Before sending an email, always confirm the recipient, subject, and body with the user."
In Generative mode, the agent may call actions unintentionally. For actions with side effects such as sending email or modifying data, clearly restrict the call conditions in the instructions. Example: "Before sending an email, always confirm the recipient, subject, and body with the user."
7. Knowledge in detail
Knowledge is the data source the agent references when generating answers. It is the core foundation for Generative Answers and determines the agent’s expertise and accuracy.
7.1 Supported Knowledge sources
| Source type | Description | Synchronization | Reference |
|---|---|---|---|
| SharePoint site | Site URL connection | Automatic every 4–6 hours | Learn |
| OneDrive | Personal/shared OneDrive files | Automatic every 4–6 hours | Learn |
| Website URL | Crawl public websites | Periodic | Learn |
| File upload | Direct upload of PDF, Word, Excel, PPT | Manual refresh | Learn |
| Dataverse | Power Platform table data | Real time | Learn |
💡 Practical tip — "document quality = agent quality"
When an agent's response quality is low, it is easy to think it is an LLM problem, but most cases are document quality problems.
When an agent's response quality is low, it is easy to think it is an LLM problem, but most cases are document quality problems.
- ✅ Recommended: documents with clear titles and tables of contents, one document = one topic, keep only current documents
- ❌ Not recommended: uploading PPT slides as-is, mixing duplicate/draft documents, documents with unclear permissions
- If you only add Instructions without Knowledge, response quality becomes highly unstable. Be sure to configure them together.
7.2 Knowledge access permissions
- SharePoint Knowledge: Inherits the user’s Entra ID permissions as-is. Documents the user cannot access are not included in responses.
- File upload: All users who can access the agent can receive the contents of the file. Be careful if it includes sensitive information.
8. Instructions writing guide
Instructions are the most important element configured on the agent’s Overview tab. They define the agent’s personality, role, and constraints in natural language, and serve as the top-level control layer for all decisions in Generative Orchestration.
8.1 Recommended structure
## 9. Role
You are an expert [business domain] agent for [organization name].
## 10. Goal
[Core purpose in 1–2 sentences]
## 11. Tone & style
- Answer in a friendly but professional tone
- Provide technical terms with easy explanations
## 12. Constraints (Do / Don't)
- ✅ Reference only information in Knowledge
- ✅ If you are not sure, say "confirmation is needed"
- ❌ Do not answer questions unrelated to internal policies
- ❌ Do not make unsupported guesses
## 13. Action-related constraints
- Before sending an email, always confirm the recipient/subject/body with the user
📖 Reference: Quickstart: Create an agent
📌 Instructions are "policy," not just "prompts"
It is okay for them to be long. They must not be ambiguous. Because Instructions are the operating policy the agent references in every conversation, they must be clear and specific. In particular, describe what not to do (Don't) as specifically as possible.
It is okay for them to be long. They must not be ambiguous. Because Instructions are the operating policy the agent references in every conversation, they must be clear and specific. In particular, describe what not to do (Don't) as specifically as possible.
14. Analytics — the key to operational improvement
The Analytics tab provides data for continuous quality improvement after an agent is deployed.
| Metric | Description | Improvement action |
|---|---|---|
| Total sessions | Number of conversations within a given period | Understand usage trends |
| Resolution Rate | Percentage of user questions successfully resolved | If low, strengthen Knowledge/Topics |
| Escalation rate | Percentage of conversations transferred to an agent | If high, improve the Fallback Topic |
| Thumbs up/down | User satisfaction feedback | Analyze low-scoring questions |
| Unanswered questions | List of questions that could not be answered | → Add topics or strengthen Knowledge |
| Usage by Knowledge source | Which sources are referenced often | Remove unnecessary sources / strengthen core sources |
📖 Reference: Analytics overview
💡 Practical tip — how to use unanswered questions
The unanswered questions list in Analytics is a "gold mine." Regularly review questions users actually ask but the agent cannot answer:
The unanswered questions list in Analytics is a "gold mine." Regularly review questions users actually ask but the agent cannot answer:
- Repeated questions → add a new topic (when a structured answer is needed)
- Questions with answers in documents → check Knowledge sources (possible parsing issue)
- Questions that cannot be answered → improve the fallback message ("For this question, contact [owner]")
15. Responsible AI and security
All generative AI capabilities in Copilot Studio are operated according to Microsoft’s Responsible AI standards.
| Area | Measure | Reference |
|---|---|---|
| Response validation | Reduce hallucinations with data grounding | Learn |
| Content moderation | Automatically filter harmful/non-compliant content | |
| Data protection | Conversation data is temporarily stored only for operational support purposes | |
| Not used for model training | Customer data is not used to train models | |
| Customer Lockbox | Customer approval required for support access |
📌 Next chapter preview
In this chapter, we covered Copilot Studio's UI structure and core concepts. In the next Chapter 3, we will walk step by step through the process of actually creating an agent — conversation design, variable management, action connection, and use of Generative Orchestration.
In this chapter, we covered Copilot Studio's UI structure and core concepts. In the next Chapter 3, we will walk step by step through the process of actually creating an agent — conversation design, variable management, action connection, and use of Generative Orchestration.