A Journey through the Microsoft 365 Copilot Architecture: From Request to Action

Lately, many of the conversations around the office coffee machine revolve around Microsoft 365 Copilot and the significant impact it will have on our daily work by integrating into desktop applications like Word, Excel, PowerPoint, and Teams. But how does it achieve this? What mechanisms drive the ‘magic’ behind Copilot’s AI? In this post, we will break down the process that our request follows from the moment we make the prompt until it is transformed into a response generated by Copilot.

Step 1: Initiating the Request

Imagine you are working in a Microsoft 365 application, such as Word or Excel. You decide to ask Copilot for help with a task, like organizing your emails or creating a presentation. This request, or “application command,” is the start of a chain of intelligent events.

Step 2: Copilot Takes the Lead

Copilot receives our command and springs into action. It acts like a conductor, ensuring that every element of the request is understood and processed appropriately. But before bringing your request to the big stage, it needs a bit of preparation.

Step 3: Enriching the Request

This is where Microsoft Graph and the semantic index come into play. Copilot uses these resources to add context to your request, such as relevant data from your emails, meetings, and documents. This “pre-processing” is essential to personalize the response to your specific needs.The semantic index, which is not a novelty of Copilot as it has been integrated into the new Microsoft search for some time, allows Copilot to offer relevant responses by providing us with a sophisticated data map, identifying important relationships and connections.At this point, it is important to emphasize that the semantic index can pose a “potential danger to our organization.” Not because of the technology itself, which is secure, but because of the employees of the companies, who often upload documentation to corporate environments without any security and privacy management. Copilot can be very responsible, but humans are not always so.Until now, it could be difficult to identify this sensitive information from a search engine, as we only found the information through the specific words we had used in the search. The security breach was “hidden” under a sea of documents. Now, with the Semantic Index, relationships and connections are created between the information, allowing us to retrieve information related to our request, which brings to light these private documents that were previously difficult to find.This topic and others related to information security and labeling are extensive, so we will address them specifically in future posts.

Step 4: Dialogue with the Language Model

With the request now enriched, Copilot introduces it into OpenAI’s LLM (Large Language Model), hosted on Azure (it is important to note that the query is made in a model hosted on our corporate Azure and therefore will not be accessible by anyone else, not even Microsoft, and our prompts and information from Graph will never be used to train the model). This model analyzes your command, understands your intention, and formulates a response. It is a process that seems simple, but behind it, there is a complex network of algorithms and machine learning.

Step 5: Processing and Validation

Once the language model has done its magic, the response is not ready yet. Copilot consults again with Microsoft Graph and the semantic index in a “post-processing” phase. Here, the response is reviewed and adjusted to ensure it aligns with the specific context of your request. The security and appropriateness of the response are also reviewed, ensuring that no information is provided that the user should not have access to and also ensuring that it is ethically correct and does not violate any privacy rights of the user or any of our colleagues.

Step 6: The Response and Actions Arrive in Your Application

Now, the response is ready to be revealed. Copilot sends it back to the Microsoft 365 application where you made the request. The response can take the form of text in a document, a series of organized emails, or a new slide in a presentation, all depending on your initial request.But Copilot not only provides answers; it also performs actions. If your request involved creating something, like a data table in Excel, Copilot will not only show you how to do it but will also perform the task for you. It is a tool designed not only to inform but to do.

Conclusion

The process from when you request something in a Microsoft 365 application to obtaining a response and seeing the actions unfold is truly incredible. Microsoft 365 Copilot goes beyond being just an assistant; it is an intelligent coworker, always ready to help us improve our skills without needing to make a greater effort. Without a doubt, the future of productivity, today.If you want to know how Artificial Intelligence can integrate and boost your company, do not hesitate to contact us.