Microsoft 365 Copilot Exceeds Expectations and Shows What Lies Beyond GPT-4
Generative AI is coming to every aspect of Microsoft's productivity apps
TLDR;
Microsoft 365 Copilot is a new large language model (LLM) based assistant behind a number of new features in the Microsoft 365 office productivity suite.
Whereas Google and others have shown LLM solutions as new features bolted onto existing applications, Copilot goes much deeper. It can analyze content and data, and is a system that conducts pre-processing of the prompt and post-processing of the response based on the Microsoft knowledge graph of information about the user and the organization they work for.
Microsoft shared videos, not live demonstrations, of several of these solutions and the features are impressive if they work as advertised.
Micrsoft said 365 Copilot is currently in use with a limited number of customers and will have wider availability after the current test period.
Microsoft today introduced Mircosoft 365 Copilot, which is positioned as your assistant for work. Several videos depicted demonstrations about how business professionals are expected to employ the new generative AI features. Satya Nadella, Microsoft’s CEO, was quoted in the announcement saying:
“With our new copilot for work, we’re giving people more agency and making technology more accessible through the most universal interface — natural language.”
The Video Demos
Below is a video from the announcement event at the time stamp where Microsoft corporate vice president Sumit Chauhan walks through how a professional can automatically create a sales proposal from notes and then convert it into the style of previous proposals and then into a PowerPoint presentation. All of these steps are executed after a simple natural language prompt.
Maybe the most impressive demonstration was Copilot for Excel (time stamp 17:47). You can ask Copilot for information about a worksheet. The demonstration begins with a question asking about three key trends indicated by the data in the Excel file. Chauhan then narrates the request to drill down on one of the trends, and Copilot creates a new worksheet with data specific to that trend.
She then asks Copilot to visualize what contributed to the sales growth, and color is added to the key data points in the worksheet. This is followed by additional analysis based on a typed question and then turning the new analysis into a chart. Copilot will “turn a sea of data into clear insights and actions,” said Chauhan.
The announcement offered some additional details:
Copilot is integrated into Microsoft 365 in two ways. It works alongside you, embedded in the Microsoft 365 apps you use every day — Word, Excel, PowerPoint, Outlook, Teams and more — to unleash creativity, unlock productivity and uplevel skills.
Today we’re also announcing an entirely new experience: Business Chat. Business Chat works across the LLM, the Microsoft 365 apps, and your data — your calendar, emails, chats, documents, meetings and contacts — to do things you’ve never been able to do before. You can give it natural language prompts like “Tell my team how we updated the product strategy,” and it will generate a status update based on the morning’s meetings, emails and chat threads.
How it Works
Microsoft didn’t just bolt GPT-4 features into each application. Instead, the company has created what it calls the Copilot System that connects three technologies: productivity applications, Microsoft Graph, and a large language model (LLM).
After you enter a natural language prompt, Copilot sends the information first through the System and then to the Microsoft Graph, which is information that the software already knows about you, your documents, and your data. This is what Microsoft calls grounding. From a practical standpoint, Copilot is determining what instructions to append to your natural language prompt before sending it to the LLM. That results in what Microsoft calls the modified prompt.
The LLM (i.e., OpenAI’s GPT-4) then processes the modified prompt and returns a result to the Copilot System for post-processing, where additional grounding takes place based on your Microsoft Graph data.
That then modifies the response, if necessary, and returns the response and/or application commands to the user. The application commands are required if you need more than just text output.
For example, if you asked PowerPoint to create a presentation with animations based on your notes from a Word document, you need the content along with the commands to PowerPoint to create slides, insert text and images, and so forth. The LLM creates the key content elements while the System modifies the generated content based on your Graph data and determines what application commands are required to generate the presentation.
Can you Use 365 Copilot Today? Nah.
Microsoft said they are launching 365 Copilot with a small number of customers. This is presumably so they can learn more about where bugs, trust, and safety issues may arise.
This has become standard practice in generative AI for big tech. OpenAI is an anomaly because they announce solutions and features when they are available to use. However, Microsoft demonstrates products or presents a canned video but then says the new features are only available to a limited group of users without saying who they are. Google announces products one-to-two days before Microsoft’s events, shows a canned video, and says the features are coming or available only to their “Trusted Testers” — which means, “Not you 😎.”
So, the videos and demonstrations are impressive, but your opportunity to validate the functionality will have to wait. It was disappointing not to see live demos, but those are likely to be available at the Microsoft Build conference scheduled for May, if not sooner. With that said, it is fair to wonder whether the new Copilot will work as well in practice as in the videos.
I suspect we saw the optimal scenarios, which may mean the user experience does not live up to marketing similar to Apple’s adventures around the launch of Siri. However, the new capabilities will offer a lot of benefits even if they fall well short of those optimal demonstrations. If they come close to matching the performance we saw today, it will dramatically change the way knowledge workers spend their time.
In the 70 years leading up to 2018, U.S. labor productivity growth averaged 2.1% annually, according to the U.S. Bureau of Labor Statistics. Between 1998 and 2005, annual productivity growth was 3.3%. However, it was only 0.8% between 2005 and 2018. Generative AI solutions like what we saw today from Microsoft may represent the best hope for productivity gains that we have seen in over 15 years.
