3.2.1. Designing Topics and Fallback Behavior
Topics are the primary routing mechanism in Copilot Studio. Each topic defines a conversational intent — what the user is trying to accomplish — and the agent's response or action for that intent. When a user sends a message, the agent's orchestration layer matches the message to the most appropriate topic and executes it.
But topics in Copilot Studio are not the rigid, keyword-matched chatbot scripts of the past. With generative orchestration enabled, the agent uses an LLM to interpret user intent dynamically, even when the user's phrasing doesn't match any predefined trigger phrases.
Topic Types:
| Type | Purpose | Example |
|---|---|---|
| Custom topics | Handle specific business intents you define | "Check order status," "Submit expense report," "Book conference room" |
| System topics | Handle platform-level events (greeting, escalation, errors) | Conversation start, fallback, escalation to human agent |
| Fallback topic | Catches messages that don't match any other topic | "I'm not sure I understand. Could you rephrase that, or would you like to speak with a human?" |
Designing Effective Topics:
- Start with user intents, not features. List what users actually ask, not what the system can do. Map user intents to topics, not the other way around.
- Keep topics focused. Each topic should handle one intent. A topic that tries to handle "order status" AND "returns" will produce confused responses.
- Design trigger phrases thoughtfully. Include 5-10 representative phrases per topic that capture different ways users express the same intent; under classic orchestration the agent triggers the topic whose trigger phrases most closely match the user's query. Under generative orchestration the trigger node shows The agent chooses and the topic's description is the most important selection factor (with its name and input/output parameters), so you don't need to predict every variation — write a short, specific description of what the topic does and doesn't do (Copilot Studio generates a default description from existing trigger phrases when you switch, which is worth revising). Make the phrases diverse — formal and casual phrasing, different word orders, abbreviations and common misspellings — rather than a single exact phrase or only formal wording. Microsoft's guidance is at least 5-10 per topic, adding more as real user phrasing comes in. If two topics misroute to each other, the usual cause is overlapping vocabulary (both use "account", "payment", "charges"): differentiate each topic's phrases with topic-specific terms and let the Multiple Topics Matched ("did you mean") system topic disambiguate close calls. Merging is for topics that genuinely do the same job; splitting into more topics just adds more overlap.
- Use topic-level variables to carry context within a topic's flow (user name, order number, selected product).
Fallback Design — The Most Underrated Topic:
The fallback topic fires when no other topic matches the user's message. Its design directly impacts user experience:
- Bad fallback: "I don't understand." (Dead end — user gives up)
- Better fallback: "I'm not sure about that. I can help with order status, returns, and product questions. Which would you like?" (Redirects to known capabilities)
- Best fallback: Uses generative AI to attempt an answer from knowledge sources, and if confidence is low, offers escalation. (Maximizes resolution while providing an escape path)
Prioritizing fallback fixes: analyze what is landing in fallback. Queries the agent should handle but has no trigger phrases for are the biggest win — add topics or phrases first. Genuinely out-of-scope queries need a helpful redirection to the right resource, not a generic message; ambiguous queries get clarifying questions. Don't delete the fallback or just pile on apology wording.
System topics and compound requests: system topics (Conversation Start, Escalate, Fallback, Multiple Topics Matched, On Error, End of Conversation and others) are built in and added to every agent. You can't delete them, only turn off ones you don't need, so customize them rather than rebuilding their behavior in custom topics. When a user asks for two things at once (for example, file a trademark and review a related contract), design for multiple intents: acknowledge both, address them in turn, and keep context across the transition — instead of merging topics or telling users to ask one thing at a time.
Exam Trap: The exam may ask what happens when generative orchestration is disabled. Without it, the agent uses classic orchestration — it selects the topic whose trigger phrases most closely match the user's message (using natural language understanding, so an exact match is not required) instead of choosing topics, tools, and knowledge by their descriptions. This is a critical architectural decision that affects agent flexibility and user experience.
Reflection Question: An IT help desk agent has 15 custom topics covering common issues. Users frequently type vague messages like "it's broken" or "nothing works." How would you design the fallback topic to handle these effectively?