AI
The Copilot Studio and Azure AI Foundry Integration Model
Copilot Studio is the SaaS starting point: deploy immediately, no GPU infrastructure to manage. Azure AI Foundry adds two things it lacks. First, full control over RAG infrastructure via Azure AI Search. Second, access to the 1900-model Foundry catalog, so teams can pick different models per task and optimize for latency, cost, or reasoning depth.
“You can also now bring in all 1900 models from the Foundry catalog.” — Chris Mullins
The tooling spans from no-code Agent Builder to the M365 Agent SDK in Python, C#, and JavaScript. Builders are not locked into SaaS or PaaS. They start in Copilot Studio and pull in Foundry resources as requirements grow.
Connecting Azure AI Search as a Knowledge Source
When SharePoint, Dataverse, and Dynamics are not enough, Azure AI Search fills the gap. It is now generally available as a knowledge source in Copilot Studio. Teams configure the embedding model, chunk size, and image extraction model in Foundry, then link the index into the agent. The demo used GPT-4o for image extraction and text-embedding-3-large for text.
The Azure AI Search index creation and agent wiring demo (08:41) ingested a 45-page Surface device service manual, diagrams included, into a blob-backed index. The agent then answered hardware questions by pulling the relevant chunks with citations.
Swapping the Response Model with a Fine-Tuned Foundry Deployment
Copilot Studio offers two model paths. Managed models come with Microsoft-handled GPU infrastructure and pre-applied responsible AI filters. Bring your own model opens the full 1900-model Foundry catalog (Claude, Llama, Gemini) but the team manages deployment and safety controls. Each agent sets its primary response model through the generative AI settings tab.
The fine-tuned model swap demo (17:15) replaced GPT-4o with a Contoso-trained Foundry deployment. The fine-tuned model surfaced the Contoso certificate refresh tool first, because training data reflected which fix actually works internally.
“So as we can see, the response is quite different.” — Uri
Per-Prompt Model Selection with Prompt Builder
Prompt Builder attaches a Foundry model per prompt, not per agent. One agent can run Phi-4 for text troubleshooting and GPT-4.1 for image analysis in the same session. When an image input is active, the platform filters out models that don’t support vision.
The Prompt Builder multi-model comparison demo (24:43) ran the same IT helpdesk prompt against Phi-4, Mistral, and Grok side-by-side. Grok had been announced in Azure AI Foundry the prior day. MK chose Phi-4 for the straightforward troubleshooting case. The feature is in public preview on Copilot Studio, Power Apps, and Power Automate.
CSX’s Production Multi-Agent Deployment
CSX, a freight rail company covering 23 states and 21,000 miles of track, built a customer-facing agent called Chessie on its ShipCSX portal. A supervisor agent authenticates users and routes to specialist agents for track and trace, frequent questions, or case creation. Cases flow automatically into Salesforce when users give a thumbs-down, so customers no longer fill out forms.
“This took I’d roughly say three months to implement” — Natalie
The build took roughly three months, the bulk of it testing before going public. CSX has a backlog of additional agents queued for ShipCSX, plus separate internal Chessie agents.
Q&A
Does Azure AI Search work with VNET subnet injection? Yes, Azure AI Search supports VNET subnet injection as a PaaS platform capability. ▶ 40:25
How do you handle authenticated knowledge bases in Azure AI Search with Copilot Studio? Azure AI Search supports authentication, and SharePoint knowledge sources pass the user’s OBO token so they only retrieve documents they have access to. ▶ 41:00
Can the agent learn from past user interactions or be fine-tuned on conversation history? Native memory is not yet in Copilot Studio, but all conversation logs are stored in the ConversationTranscript table, and teams can build a memory tool that extracts facts into an Azure AI Search index for recall. ▶ 41:38
How do you prevent the agent from adding extra JSON wrapper layers around a structured output? Prompt Builder lets you define an explicit output schema; for direct API use, the OpenAI JSON Schema return type enforces the schema and throws errors on violations. ▶ 47:37
Notable Quotes
Copilot Studio is the place that we recommend people start. Chris Mullins · ▶ 2:54
So as we can see, the response is quite different. Uri · ▶ 20:32
This took I’d roughly say three months to implement Natalie · ▶ 39:09
It’s definitely cost efficient. Natalie · ▶ 37:40
Key Takeaways
- Azure AI Search is now GA in Copilot Studio, giving teams full control over embedding model, chunk size, and multimodal extraction.
- Agents can use different Foundry models per prompt, letting builders optimize each task for cost, speed, or reasoning depth.
- CSX built a production multi-agent system serving 23 states in roughly three months using Copilot Studio and Foundry together.