A Walkthrough of Microsoft Copilot for Azure. What It Is, How It Works!

The speaker introduces Azure Co-Pilot, a technology that utilizes generative AI and large language models to interact with users through natural language. The video aims to explain how the technology works, emphasizing the importance of understanding its underlying mechanisms to build trust and adoption. The speaker explains that Azure Co-Pilot is based on GPT-4 from OpenAI, which is trained on vast amounts of data to predict the next word in a sequence, creating a natural interactive experience. The video also discusses how Microsoft has adopted this technology, making it read-only and not fine-tuning the models, but rather adapting the interactions through prompts and additional information.

The speaker delves into how AI interactions work in the context of Azure Co-Pilot. He explains that the AI requires more than just a prompt to be useful; it needs contextual information. The speaker outlines how Azure Co-Pilot interacts with various Azure services, such as Microsoft Docs, Azure Resource Manager, Azure Resource Graph, and others, to gather necessary data. He emphasizes that Microsoft is cautious about exposing functionalities that can modify or delete resources, ensuring that the right guardrails and confirmations are in place. The speaker also explains that the AI model does not directly access any resources; it operates through the Azure Co-Pilot, which enforces responsible AI principles.

The speaker discusses the safety and compliance aspects of using Azure Co-Pilot. He explains that the Co-Pilot operates within the user's permissions, meaning it can only perform actions the user is authorized to do. The speaker reassures that the Co-Pilot does not have its own service principle with full permissions and that it cannot bypass any policies or guardrails. He also highlights that the Co-Pilot is not a back door to resources and that it cannot do anything the user couldn't already do through the portal. The speaker compares the introduction of Co-Pilot to the initial concerns around PowerShell, emphasizing that it is about efficiency, not bypassing restrictions.

The speaker provides a demonstration of Azure Co-Pilot's functionality. He shows how Co-Pilot can change the theme of the Azure portal and how it can generate CLI scripts. The speaker also demonstrates how Co-Pilot interacts with Azure services, such as resource graph queries and understanding KQL (Kusto Query Language). He highlights that Co-Pilot can help with tasks like listing unused public IPs and provides suggestions on how to deal with them. The speaker emphasizes that Co-Pilot can only do what the user can do, and it makes the user's job more efficient by guiding them through tasks and providing helpful suggestions.

The speaker discusses the interaction patterns of Azure Co-Pilot and its limitations. He explains that Co-Pilot can perform multiple iterations to gather the necessary information and provide the correct response. The speaker also addresses the limitations around the number of interactions per chat and the number of chats per day, noting that these limitations are due to the computational power required for training and inferencing. He mentions that there are no pricing details yet and that these will be announced once the technology moves out of private preview and becomes generally available.

The speaker addresses common questions about Azure Co-Pilot, such as controlling which subscriptions can use Co-Pilot and preventing Co-Pilot from making changes. He clarifies that Co-Pilot operates at a tenant level and cannot be turned on or off at a subscription level. The speaker advises that to prevent Co-Pilot from making changes, organizations should manage user permissions correctly, emphasizing a zero-trust and least-privilege approach. He suggests that if users do not have the permissions to make changes in the portal, Co-Pilot will also not be able to make changes. The speaker concludes that Co-Pilot is designed to help users do their jobs better and that it operates within the same policies, APIs, and restrictions as the user.

The speaker concludes the video by reiterating that Azure Co-Pilot is designed to enhance the user's efficiency and effectiveness without enabling them to do anything they couldn't already do. He emphasizes that Co-Pilot operates within the user's permissions and that it is not a separate entity with its own permissions. The speaker also highlights that Co-Pilot is not used to train the model, which is a read-only copy from OpenAI, and that it operates within Microsoft's security and regulatory boundaries. He assures viewers that Co-Pilot is there to help them do their jobs better and that all its interactions are guided by responsible AI principles.