Microsoft M365 Copilot

Microsoft M365 Copilot uses generative AI described in Develop generative AI apps in Azure - Generative AI in Azure AI Solutions.

Functions include:

• Accessing your data for context and prompt
• Agents
• History of previous conversations
• Customization using instructions

Architecture

                     +-------------------------------+
User --> Copilot --> |   Large Language Model (LLM)  |
           |         +-------------------------------+
           |              modified prompts
           |
           |--> Microsoft Graph for messages, context, files
           |        |
           |        |--> Dataverse, Power Platform services
           |        |--> Indexed vector databases related to LLM model version chosen
           |        |--> Email, Messaging, Cloud files, Collaboration websites
           |
           |--> Plugins, Bing search
           |<-- Organization controls, constraints

See Microsoft Copilot Architecture and for Microsoft Graph functions see Graph Explorer for example functionality

Cognitive Services

Q: Related to agents, cognitive services: Saw M365 Copilot can do NLP, vision, and information extraction, for example, OCR of written notes and heading and list recognition. As an agent, does Copilot call the formerly named cognitive services behind the scenes for users and process results? What model is Copilot using? Was thinking in the context of agent design, it will be helpful to use the services like tools specific to language, vision, documents tasks. For example, at work there is a sequential process with OCR of forms and field extraction using text parsing currently.

Copilot is like a complex agent and has access to many tools. As of 2025, Copilot uses OpenAI models hosted in Azure. Priorities in choosing the model is safety (low vulnerability to jail breaking) and accuracy.