created with freepik.com

Introduction

Today, I’m going to try to explain in detail how Copilot for Microsoft 365 works. I’ve been reading a lot of articles about the underlying architecture and processes, and here I’m going to give you a simplified summary of what I understand. Last week, I heard the short explanation of Copilot for Microsoft 365 being “ChatGPT on steroids.” I don’t think that’s completely right. Copilot for Microsoft 365 is a tool that integrates into the office suite. It’s built for specific use cases and can construct answers from information on the internet and your company’s data.

The Architecture

The language model Copilot for Microsoft 365 uses is based on OpenAI’s GPT (Generative Pre-trained Transformer) model. Copilot for Microsoft 365 takes the functionality of GPT to the next level by integrating it with the user’s data in Microsoft Graph. This ensures that the responses are personalized and relevant.

From Microsoft Learn

Copilot follows the flow shown in the above picture. First, Copilot gets a prompt from a user in an app like Microsoft Word. Then, Copilot pre-processes the input through a process called grounding to make the prompt more specific. In the grounding process, Copilot accesses data through Microsoft Graph. When enabled, it can also search the web or other services for grounding. You can also add more sources to Copilot in Teams if you want to. The grounding process is really important because it’s what makes Copilot better than traditional LLMs. Microsoft isn’t fine-tuning the LLM with your data, but the prompt that Copilot sends to the LLM is enriched with your personal data and intelligently retrieved data from the internet. Because of the grounding process and the data being used, Copilot is less likely to write wrong answers that are known in the world of AI as hallucinations.

The enriched prompt, which includes data from Microsoft Graph and possibly external sources like the internet, is sent to the Large Language Model (LLM) for processing. As you can see in the architecture picture, the LLM is a central service and not a dedicated instance for each customer. Each request sent to the LLM is a complete, stateless transaction that includes all relevant information. Once the LLM has processed the information, it doesn’t store it and doesn’t use it to train the model. As you might have noticed, your old prompts are still available. They’re stored inside your Exchange Online mailbox, just like your Teams chat messages.

Finally, Copilot takes the response generated by the LLM and processes it further, making additional calls to Microsoft Graph, doing responsible AI checks, security, compliance, and privacy reviews. In addition to the Purview compliance controls, the Copilot service includes its own guardrails that can’t be customized. These guardrails prevent harmful or unethical content from being shown to the user. The response is then returned to the app and shown to the user.

Integration

Copilot for Microsoft 365 is all about making your life easier by helping you get things done faster. It’s like having an assistant in your Microsoft 365 apps, like Word, Excel, PowerPoint, Outlook, Teams, and more. All you have to do is ask it a question or give it a command in natural language, and it’ll do the rest.

Copilot interprets the prompt, accesses the relevant data from Microsoft Graph, and uses its LLM to generate a response or perform a task. The response is then presented within the app, often with the option to edit or refine the output.

If you ask Copilot to summarize a document, it’ll analyze the content of the document stored in SharePoint or OneDrive, extract the key points using its LLM, and present a concise summary. Similarly, if you ask Copilot to create an email draft, it will generate the text based on the context you provide and your previous communications.

Just a heads-up: Copilot’s responses depend on what you tell it. The more specific you are in your prompts, the more accurate and helpful the response will be. It’s a good idea to try out different ways to write a prompt.

Conclusion

To sum up, Copilot for Microsoft 365 is a pretty powerful tool that combines the advanced capabilities of GPT with Microsoft’s extensive data ecosystem to provide a personalised and efficient productivity assistant within the Microsoft 365 suite. I just want to reiterate that Copilot is not going to fully automate your work, but it can help you in some areas where you might be struggling. You’ll have to figure out what those areas are and how you can benefit most from the technology.

I hope you enjoyed this article. Stay tuned for more!

Additional Links

Here are some extra links to help you get a better understanding of how Copilot works:

How Copilot for Microsoft 365 Works | Practical365

Microsoft Copilot for Microsoft 365 overview | Microsoft Learn

Microsoft Copilot Architecture – Explained (redresscompliance.com)

Copilot for Microsoft 365 : Architecture and Key Concepts – Microsoft Community Hub

Schreibe einen Kommentar

Deine E-Mail-Adresse wird nicht veröffentlicht. Erforderliche Felder sind mit * markiert

Consent Management Platform von Real Cookie Banner