Microsoft Copilot Agents Builder: Topic-Based vs Generative Modes

You want to build a Copilot agent that answers questions accurately for your team but are unsure whether to use topic-based or generative mode. Topic-based mode restricts the agent to a fixed set of predefined answers while generative mode allows the agent to compose responses from scratch using large language models. This article explains the differences between the two modes, how each handles grounding data, and which scenarios favor one over the other. You will learn the technical trade-offs and see a direct comparison so you can choose the right mode for your Copilot agent.

Topic-Based Mode and Generative Mode Defined

The Copilot Agents Builder in Copilot Studio lets you create custom agents that can be embedded in Microsoft Teams, websites, or other channels. The core architectural decision is the mode you select during agent creation. Topic-based mode, inherited from the legacy Power Virtual Agents, forces the agent to follow a rigid conversation flow. You define topics manually, each topic contains a trigger phrase and a fixed response. The agent matches user input to a trigger phrase and replies with the exact answer you wrote. It never generates new text or interprets beyond the defined topics. This mode is predictable and auditable because every possible answer is prewritten.

Generative mode, introduced in 2024, replaces the topic-based flow with a large language model. The agent receives a system prompt that describes its role, boundaries, and tone. It then uses uploaded documents, SharePoint files, or web content as grounding data. When a user asks a question, the agent retrieves relevant chunks from the grounding data and passes them to the language model, which composes a natural-language answer. The agent can also fall back to general knowledge if the grounding data does not contain an answer, depending on your configuration. This mode produces richer, more flexible responses but introduces variability because the model may phrase answers differently each time.

How Grounding Works in Each Mode

Grounding refers to the data the agent uses to answer questions. In topic-based mode, grounding is explicit and finite: every topic you create is a piece of grounding. There is no external data source. In generative mode, grounding comes from knowledge sources you attach to the agent. You can upload PDFs, Word documents, PowerPoint files, Excel spreadsheets, and connect to SharePoint folders or public websites. The agent indexes these sources and performs retrieval-augmented generation, or RAG. The model only uses the retrieved content to compose its answer. If the retrieved content does not contain the answer, the agent can either say it does not know or attempt to answer from its training data, controlled by the “Use general knowledge” toggle in Copilot Studio.

When Each Mode Was Introduced

Topic-based mode has been available since the Power Virtual Agents platform launched in 2020. Generative mode became available in Copilot Studio in early 2024. Microsoft recommends generative mode for new agents unless you need strict compliance or auditability that only predefined answers can provide.

Steps to Create an Agent and Choose a Mode

You must select the mode when you first create an agent. You cannot change the mode after creation. Follow these steps to set up a new agent in Copilot Studio.

1. Open Copilot Studio
   Go to copilotstudio.microsoft.com and sign in with a work or school account that has a Copilot Studio license. Click the Agents tab in the left navigation.
2. Start a new agent
   Click Create agent on the top toolbar. A dialog appears asking you to choose the agent type. Select Custom agent.
3. Choose the mode
   In the “Choose how your agent answers” section, you see two radio buttons: Topic-based and Generative. Select one. A short description appears next to each option. If you select generative, you must also provide a name and a description for the agent.
4. Set up knowledge sources for generative mode
   If you chose generative mode, the next step shows the Knowledge tab. Click Add knowledge and choose from SharePoint, files, web pages, or Dataverse. For files, you can upload up to 20 documents with a total size limit of 100 MB. For SharePoint, select a site and folder. The agent indexes the content immediately.
5. Create topics for topic-based mode
   If you chose topic-based mode, go to the Topics tab. Click Add topic and enter a name. Under Trigger phrases, type at least five example phrases a user might say. Under Response, type the exact answer the agent should return. You can also add branching logic with conditions and variables.
6. Test and publish the agent
   Click the Test button in the bottom-left corner to open the test pane. Type sample questions and verify the agent responds as expected. When ready, click Publish and then Publish agent to make it available in Teams, a website, or a custom channel.

Common Mistakes and Limitations

Selecting the Wrong Mode After Creation

You cannot switch between topic-based and generative modes after the agent is created. If you need to change modes, you must create a new agent. To avoid this, review your requirements before creation. If you need flexible responses and can accept some variability, choose generative. If you need fully controlled answers, choose topic-based.

Generative Mode Producing Incorrect Answers

If the agent answers incorrectly, the grounding data may be incomplete or the “Use general knowledge” setting may be enabled. In Copilot Studio, open the agent, go to Settings > Generative AI, and disable Use general knowledge. This forces the agent to answer only from your uploaded content. Also verify that the uploaded documents contain the exact information users need. The agent cannot infer or combine facts across documents unless the same concept appears in multiple sources.

Topic-Based Mode Cannot Handle Unseen Questions

If a user asks a question that does not match any trigger phrase, the agent responds with the default fallback topic. You must manually add topics for every question you expect. This limitation makes topic-based agents brittle for open-ended support scenarios. If you frequently see “No match” responses, consider switching to generative mode in a new agent.

Generative Mode Latency

Generative responses take longer than topic-based responses because the agent must perform retrieval and then generate text. Typical response times range from 2 to 6 seconds. Topic-based responses are instant because they return a static string. If your users need sub-second responses, topic-based mode is the better choice.

Topic-Based vs Generative Modes: Key Differences

Item	Topic-Based Mode	Generative Mode
Response method	Predefined static text	AI-generated from grounding data
Grounding data	User-created topics only	Uploaded files, SharePoint, web pages
Flexibility	Low – only answers from topics	High – can rephrase and combine data
Auditability	Full – every answer is known in advance	Partial – answers vary by run
Setup effort	High – must write all topics manually	Medium – upload documents and write a prompt
Response time	Instant	2–6 seconds
Fallback handling	Fixed “No match” topic	Configurable: say “I don’t know” or use general knowledge
Best for	FAQs with strict compliance	Open-ended Q&A, knowledge base agents

You now understand the difference between topic-based and generative modes in the Copilot Agents Builder. Topic-based mode gives you full control over every answer but requires manual topic creation. Generative mode lets the AI compose answers from your documents, saving setup time while introducing some variability. Start by evaluating your compliance requirements and the breadth of questions users will ask. If you need absolute predictability, build a topic-based agent. If you need a flexible agent that can handle diverse questions, build a generative agent and disable general knowledge to keep answers grounded. After publishing, monitor the analytics in Copilot Studio to see which questions the agent answers correctly and adjust your grounding data or topics accordingly.