Copilot Studio: Topics, Flows & Prompt Actions

How Copilot Studio agents think

NLP vs CLU vs generative AI orchestration

Three language understanding approaches in Copilot Studio

Feature	How It Works	Best For	Trade-offs
Standard NLP	Trigger phrases on each topic — keyword matching with basic natural language understanding	Simple, predictable scenarios with well-defined user intents. Internal tools where users know the vocabulary	Fast to set up, deterministic. But brittle — struggles with unexpected phrasing or synonyms
Azure CLU	Trained intent classifier and entity extractor — you provide labelled examples, CLU learns the patterns	Domain-specific language where standard NLP fails. Regulated industries needing auditable intent classification	Higher accuracy for trained intents. Requires labelled training data and ongoing model updates as language evolves
Generative AI orchestration	LLM dynamically understands intent and either routes to a topic or generates a response from knowledge sources	Open-ended conversations, broad knowledge domains, scenarios where users phrase requests unpredictably	Most flexible and handles novel phrasing. But less predictable — needs guardrails, content safety, and testing for hallucination

💡 Exam tip: the orchestration decision

The exam tests whether you can recommend the right approach for a given scenario:

• “Users ask the same 15 questions in predictable ways” → Standard NLP with trigger phrases
• “Users describe problems in domain-specific jargon that varies by region” → CLU with trained intents
• “Users ask open-ended questions across a large knowledge base” → Generative AI orchestration
• “The solution must provide auditable intent classification for compliance” → CLU (deterministic, logged)
• “The agent needs to handle both structured workflows AND open Q&A” → Hybrid — generative orchestration for routing, with topics for structured flows

The hybrid approach is often the right answer on the exam. Generative orchestration handles the “understanding” layer, and structured topics handle the “action” layer.

Topic types in Copilot Studio

Topics are the building blocks of a Copilot Studio agent. Each topic handles a specific user intent or system event.

Topic Type	Purpose	Example
Custom topics	Handle specific user intents you define — the core of your agent design	”Track my order,” “Reset my password,” “Request a refund”
System topics	Built-in topics that handle common events (greeting, goodbye, escalation, error)	Greeting topic fires when a user starts a conversation
Fallback topic	Catches any message that no other topic matches — your safety net	”I’m not sure I understand. Would you like to speak with a human?”

Fallback design is critical. A poorly designed fallback frustrates users. A well-designed fallback either:

1. Routes to generative answers — “I don’t have a specific workflow for that, but let me check our knowledge base…”
2. Asks a clarifying question — “I can help with orders, returns, or account issues. Which one?”
3. Escalates gracefully — “Let me connect you with someone who can help with that.”

Agent flows: multi-step workflows

Agent flows are the multi-step workflows inside Copilot Studio topics. They define what the agent does after it understands the user’s intent.

Design elements of an agent flow:

1. Trigger — what starts the flow (user message, event, schedule)
2. Conditions — branching logic based on data or user responses
3. Actions — calls to connectors, Power Automate flows, or external APIs
4. Variables — store and pass data between steps
5. Messages — responses sent back to the user at each stage
6. Escalation — hand off to a human agent when needed

💡 Scenario: Natalie designs a hybrid agent for a client's support portal

Natalie Torres (Cloudbridge Partners) designs an agent for a telecom client’s customer support portal. The agent needs to handle both structured workflows (plan changes, billing inquiries) and open-ended product questions.

Design decisions:

Orchestration: Generative AI orchestration as the primary router — it understands user intent from natural language and either triggers a structured topic or generates an answer from the knowledge base.

Structured topics (agent flows):

• “Change my plan” → Authenticate user → Show current plan → Present options → Confirm change → Call billing API → Confirm to user
• “Dispute a charge” → Authenticate → Pull billing history → Collect dispute details → Create case in D365 → Provide case number

Generative fallback: For open-ended questions (“Does 5G work in rural areas?”), the orchestrator searches the knowledge base and generates a contextual response.

Fallback topic: If confidence is low on both structured and generative paths: “I want to make sure I give you the right answer. Let me connect you with a specialist.”

Zoe Park (PM) tracks the design: 40% of interactions hit structured topics, 50% are handled by generative answers, and 10% reach the fallback escalation.

Prompt actions

A prompt action is a custom AI-powered step within a topic that calls an LLM with specific instructions. It lets you inject AI reasoning at any point in an agent flow.

When to use prompt actions:

• Summarise — condense a long document or conversation into key points
• Classify — categorise user input (sentiment, urgency, product category)
• Extract — pull structured data from unstructured text (dates, amounts, entities)
• Generate — create personalised responses, recommendations, or content
• Transform — rewrite text for a different audience or format

Design considerations:

• Prompt actions use tokens — design prompts to be concise to control cost
• Output should be validated before passing to downstream steps
• Include guardrail instructions in the prompt (“respond only about our products,” “do not provide medical advice”)

Prompt Action Step	Example Prompt	Output Used For
Classify urgency	”Classify this support message as Low, Medium, or High urgency based on these criteria…”	Routing to the right queue
Summarise case history	”Summarise this customer’s case history in 3 bullet points for the support agent”	Agent handoff context
Generate recommendation	”Based on this customer’s usage data, recommend the best plan from our current offerings”	Personalised upsell suggestion
Extract entities	”Extract the product name, order number, and issue description from this message”	Populating a case form

Flashcards

Question

What is the fallback topic in Copilot Studio and why is its design critical?

Click or press Enter to reveal answer

Answer

The fallback topic fires when no other topic matches the user's message. It's critical because it handles every unexpected input — a bad fallback frustrates users and increases abandonment. Design it to either route to generative answers, ask clarifying questions, or escalate gracefully to a human.

Click to flip back

Question

When should you use Azure CLU instead of standard NLP in Copilot Studio?

Click or press Enter to reveal answer

Answer

Use CLU when users speak in domain-specific jargon, use varied phrasing, or when you need auditable intent classification for compliance. CLU uses trained models with labelled examples — more accurate than trigger-phrase matching but requires training data and maintenance.

Click to flip back

Question

What is a prompt action in Copilot Studio?

Click or press Enter to reveal answer

Answer

A custom AI-powered step within a topic that calls an LLM with specific instructions. Used to summarise, classify, extract, generate, or transform text at any point in an agent flow. Each prompt action consumes tokens — design prompts to be concise.

Click to flip back

Question

What is the recommended design for a hybrid agent that handles both structured workflows and open Q&A?

Click or press Enter to reveal answer

Answer

Use generative AI orchestration as the primary router to understand user intent. Route to structured topics (agent flows) for well-defined workflows. Fall back to generative answers from a knowledge base for open-ended questions. Use the fallback topic for low-confidence scenarios that need human escalation.

Click to flip back

Knowledge check

Knowledge Check

Natalie's telecom client has a customer support agent built with standard NLP trigger phrases. Customers frequently complain that the agent responds with 'I don't understand' — even when asking valid questions using slightly different words than the trigger phrases. What should Natalie recommend?

AAdd more trigger phrases to every topic to cover all possible wordingsBReplace standard NLP with generative AI orchestration so the agent dynamically understands user intentCRemove the fallback topic so users are always routed to a human agentDSwitch to Azure CLU to train a custom intent model, or enable generative orchestration — depending on whether the domain needs deterministic classification or flexible understanding

Check Answer

Knowledge Check

Ravi is designing an agent flow for a client's return processing. The flow needs to: authenticate the customer, look up the order, determine if the item is eligible for return based on policy, and either process the return or explain why it's ineligible. At which step should Ravi use a prompt action?

AAuthentication — use AI to verify the customer's identityBOrder lookup — use AI to search the order databaseCEligibility determination — use AI to interpret the return policy against the specific order detailsDAll steps should use prompt actions for maximum AI capability

Check Answer

Knowledge Check

A Copilot Studio agent uses generative AI orchestration. A user sends a message that could match either a structured 'billing inquiry' topic or a generative knowledge-base answer. What determines which path the agent takes?

AThe orchestrator evaluates confidence scores — high confidence for a topic routes there; otherwise, the generative path handles itBThe agent always prefers structured topics over generative answersCThe user must choose between asking about billing or asking a general questionDTopics and generative answers cannot coexist — you must choose one approach

Check Answer

🎬 Video coming soon

Next up: Power Apps, WAF & Data Processing — applying the Well-Architected Framework to intelligent workloads, embedding AI in Power Apps, and designing data processing pipelines for grounding.