Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
[This article is prerelease documentation and is subject to change.]
In Copilot Studio streaming mode, a single generative AI model powers the entire agent. This approach eliminates the need for traditional intent trees, rigid dialog flows, or separate orchestration layers.
Instead of stitching together multiple systems, the model acts as the brain of the agent. It dynamically decides how to respond and what actions to take in real time.
Important
- This is a preview feature.
- Preview features aren’t meant for production use and might have restricted functionality. These features are subject to supplemental terms of use, and are available before an official release so that customers can get early access and provide feedback.
What the generative model actually does
The generative AI model handles the complete conversational lifecycle:
Understanding the user’s request (intent and entities)
Deciding the next best action, such as:
Answering from knowledge sources
Calling tools, APIs, or MCP integrations
Asking for clarifying follow-up questions
Executing those actions
Synthesizing a natural, conversational response (voice or text)
This approach represents Microsoft’s generative orchestration model. It replaces:
Traditional NLU intent classification (Classic orchestration mode)
Hard-coded dialog routing
Deterministic decision trees
The result is a much more flexible, human-like interaction model.
When to use knowledge, APIs, or MCP
At a high level, the difference between knowledge, APIs, and MCP depends on the type of answer or action needed. Use knowledge when the agent needs to explain policies, FAQs, disclaimers, or other guidance based on information that rarely changes. Use APIs when the agent needs real-time, customer-specific data, such as checking an order status or retrieving account details from a system of record. Use MCP when you need scale and structure. It provides a standardized way for the model to interact with multiple tools and systems in a consistent, governed manner.
In simple terms: knowledge helps the agent talk smarter, APIs help it act on live data, and MCP helps it do both reliably across complex enterprise environments.
| Approach | When to use | Key characteristics | Examples | How it works |
|---|---|---|---|---|
| Knowledge (static content) | Answer is static or rarely changes. No personalization required. No system validation needed. | Fast and low cost. No backend dependency. Ideal for FAQs and general info. | What are your store hours? Do you offer refunds on sale items? How does your subscription plan work? | Model retrieves from knowledge base. Generates direct response without external calls |
| APIs (real-time data) | Need real-time transactional data. Request is user-specific. Requires authoritative system response. | Dynamic and personalized. System of record driven. Ensures accuracy and freshness | What is my order status? Do I have any upcoming appointments? Has my refund been processed? What’s my current balance? | Model detects need for live data. Calls API (Power Automate, connector, HTTP endpoint). Receives structured data. Converts to natural language response. |
| MCP (tool orchestration layer) | Need standardized, reusable integrations. Multiple systems/tools must work together. Require scalable, governed access for LLMs. | Acts as a contract layer between model and tools. Enables orchestration across systems. Safe and more scalable for enterprise use. | Pull CRM data. Check billing systems. Update ticketing systems. Cancel my subscription and refund the last charge if eligible | Model selects MCP tools. MCP orchestrates multiple backend systems. Executes workflows safely. Returns structured results for response generation |