Microsoft 365 Copilot and Copilot Studio Demo: Microsoft Build 2025

Key Concepts:

• Microsoft 365 Copilot app
• Copilot Studio
• Copilot Tuning
• Agents (Researcher Agent, RFP Response Agent, Compliance Review Agent, Contract Builder Agent)
• Connectors (GitHub Connector)
• Microsoft Graph APIs
• AI Foundry
• Multi-Agent Orchestration
• Data Labeling
• Fine-tuning
• GPT-4o
• Entra ID

1. M365 Copilot App: A Five-in-One Hub

• The M365 Copilot app is presented as a central hub for work, encompassing Chat, Search, Agents, Notebooks, and Create functionalities.
• Example: The Researcher Agent is showcased, demonstrating Copilot's ability to reason over work data and external sources like GitHub via connectors.
• Process: The user initiates a request (e.g., analyzing performance issues), the Agent asks clarifying questions, and then leverages the OpenAI o3 reasoning model to gather information from various sources.
• Technical Detail: The Researcher Agent uses a chain of thought process, providing citations for its findings.

2. Building Connectors with Visual Studio and Microsoft Graph APIs

• The process of building a connector to integrate external data sources (e.g., GitHub) is outlined.
• Step-by-Step:
  1. Create a new connector project in Visual Studio using the M365 Agents Toolkit. This provides fully scaffolded code.
  2. Use Microsoft Graph APIs to establish a connection.
  3. Define a schema for the data to be indexed (e.g., GitHub issues).
  4. Use the GitHub API to fetch issues.
  5. Ingest the issues into Microsoft Graph, making them accessible to Copilot.
• Technical Detail: The demonstration shows the indexing process and confirms the GitHub connection and backlog items.

3. Copilot Studio: Building Agents with Low-Code

• Copilot Studio enables the creation of agents with minimal coding.
• Example: An RFP (Request for Proposal) response agent is presented, capable of generating proposals and posting them using its own Entra ID.
• Process:
  1. Describe the agent's purpose and provide instructions within Copilot Studio.
  2. Select the desired response model (GPT-4o by default, with options from AI Foundry).
  3. Specify the knowledge sources for grounding the agent.
  4. Add triggers to initiate the agent's work (e.g., a new RFP arriving in the inbox).
  5. Connect the agent to data sources (e.g., Dynamics MCP Server for SKU and pricing data, third-party servers like Docusign, or custom MCP servers like SAP for customer account data).

4. Multi-Agent Orchestration

• Multi-agent orchestration allows agents to collaborate on complex tasks.
• Example: The RFP response agent is connected to a compliance review agent to ensure no red flags exist before proceeding.
• Process: The RFP agent connects to the compliance agent, which performs a compliance check. If the check passes, control returns to the RFP agent to continue the proposal process.

5. Copilot Tuning: Low-Code Model Fine-Tuning

• Copilot Tuning offers a low-code approach to fine-tuning models, traditionally a task requiring data scientists.
• Example: A contract writing model is fine-tuned using example documents.
• Process:
  1. Create a new model, providing a name, description, and task type (e.g., document generation).
  2. Add a knowledge source (e.g., a contract database from SharePoint).
  3. Specify access permissions for the knowledge source (e.g., Contract Team and Procurement Team).
  4. Ensure Copilot aligns access to knowledge sources in the tuned model.
  5. Complete data labeling by subject matter experts.
  6. Complete the training for the fine-tuned model.
  7. Publish the model.
• Technical Detail: The fine-tuned model allows the Contract Builder agent to generate contracts using the company's specific language, terms, conditions, structure, and format.

6. Using Fine-Tuned Models in the M365 Copilot App

• The fine-tuned model can be accessed and used within the M365 Copilot app.
• Process: Users can select "Create Agent," choose a task-specific agent, and then select the fine-tuned model (e.g., the Contract Builder model).

7. Synthesis/Conclusion

• The presentation highlights how developers can leverage Microsoft 365 Copilot app, Copilot Studio, and Copilot Tuning to scale productivity solutions.
• The tools empower individuals closest to the business to reimagine their workflows with AI, enabling them to automate tasks, access relevant data, and generate tailored content.
• The emphasis is on low-code approaches, making AI capabilities accessible to a wider range of users and reducing the reliance on specialized data science teams.