Excel is intimidating to many people, as it is used by millions daily, but only a few consider themselves experts in all of its features. In 2015, I joined Microsoft's Excel design team and embarked on a project aimed at making data less daunting. My objective was to design a solution where AI could work alongside users, empowering them to engage more deeply with their data without needing to learn all the intricacies of Excel. At the time of my joining, there was no clear vision or roadmap to address the data literacy problem. Starting with research and developing high-level design concepts, I pitched the project successfully, securing funding. Subsequently, I took charge of leading and owning all aspects of UX decisions, collaborating with over 25 PMs, engineers, data scientists, and design researchers. The project was shipped to millions of users worldwide and reached a monthly active user base of 1 million in 2020. Over time, Excel Ideas has evolved and influenced the recent announcement of Microsoft Copilot.

I will now break down my 3-year journey on this project into three phases:

Phase 1 - Creating a Product VisionPhase 2 - Design Iteration and Shipping to Production
Phase 3 - Expanding the Scope & Lead Coherence in Office

As the sole product designer and founding member of the project, I collaborated with an engineering manager, a PM lead, and a researcher in the early phase of the journey. For the initial 4 months, we conducted extensive user research and visited numerous users at their workplaces to see how they work with data. Our findings revealed that thousands and millions of people rely on Excel for making business decisions, with 80% of their time dedicated to three key steps highlighted in green below. In practice, there is an abundance of data but a scarcity of data skills. Excel's existing design expected users to learn and adapt to our software in order to answer simple data-related questions, such as "How does my average sales for coffee beans this year compare to last year?" However, users often encountered results that didn't align with their assumptions, turning the process into an iterative journey.

Once I identified the problem and the opportunity, I led a one-week design sprint with all project members and other designers on the Excel team to explore various solutions. There were numerous ideas surrounding the utilization of charts, visualization, and an artificial intelligence project that the Microsoft Research lab was working on at the time. I quickly put together some low-fidelity mockups and collaborated with an in-house video production team to create a video illustrating the project vision that automatically analyzing users' data and highlights the key trend and insights for the users

It is important to note that the mockups in the video were low fidelity and were created solely for the purpose of storytelling. However, the narrative, exemplified by inspiring users like Mindy, successfully secured funding for our project with a team of engineers to bring it to life. Intelligent data analysis also emerged as one of the 4 core initiatives for Excel and Office for that fiscal year, extending into the present day.

Based on the design sprint work and user research completed by the Excel team, I discovered that people tend to comprehend data better when presented in chart form. Collaborating with data scientists during Phase 1, we focused on the insights that users valued the most from our previous research, such as correlations, outliers, trends, majorities, and more.

Before generating insights for any data, Excel needed to understand the data's meaning. For instance, "Ratings" typically represented a numerical value, "Brand" was categorical, and "Year" appeared as a number but technically served as a category name and shouldn't be summed together. While the data scientists worked on enhancing the accuracy of their machine learning models for predicting data types, I dedicated considerable time to studying data visualization best practices. I engaged with data visualization experts within the company, delved into books like Edward Tufte's works and "Storytelling with Data" by Cole Knaflic, and conducted interviews with a Bloomberg columnist. My goal was to create charts that were both simple and allowed users to swiftly grasp key insights, whether related to correlations, trends, or identifying outliers. I opted for the basic chart types such as line, bar, scatter, which are native to Excel, enabling users to learn and create their own charts later on. Additionally, I utilized color highlighting to draw attention to vital information.

"Mobile first" was a prominent design principle. However, considering that the majority of Excel users operated it on Windows machines, we decided to prioritize the project's development on Windows initially. Consequently, we built it as a web app add-in using React, seamlessly integrated within Excel on Windows. Subsequently, we successfully embedded the same web add-in into Excel for Mac and Excel Online on the web. This presented us with a visual design challenge: the app had to seamlessly blend with the surrounding UI across all four color themes on Windows, two color themes on Mac, and in any web browser.

Even though we did not have plans to ship on iOS and Android at that time, I dedicated some time to studying the mobile design patterns used in the Excel iOS app. I specifically explored 5 different entry points for Ideas (formerly known as Insights) and 4 interaction models. All of these explorations were based on the existing UI patterns in the Excel iOS app.

There were several problems we needed to address before shipping, including localization, accessibility, and, most importantly, establishing metrics to measure success. When Ideas was initially released to the public, it was fascinating to observe the live telemetry data as users began exploring the feature. We defined our measure of success with the "seen, tried, kept" framework, consisting of 3 steps in the funnel: the number of people who saw the feature, the number of users who tried it out, and the number of charts that were retained in the spreadsheet. I heavily relied on the telemetry data on a daily basis to brainstorm ways to enhance the user experience.

For example, from telemetry data, we noticed that user encountering an error message was quite common, often due to issues such as a faulty internet connection or a dataset being too large. This resulted in a negative experience and often became a dead end for the user. Upon analyzing the data, I discovered that 38% of the error messages occurred when users clicked the "Ideas" button on a blank sheet with no content. The previous version of the error message simply indicated that it was a blank sheet and advised users to try a different dataset. To improve the experience, I implemented a better error message that included a call-to-action button offering sample data and a visual example. As a result of this updated UX for the error message, we observed a significant increase in the telemetry metrics for "seen, tried, kept."

Another example derived from analyzing the telemetry data was that users often opened the Insights feature but did not frequently engage with the "insert chart" button. This created a significant gap between the "seen" and "tried" stages. I understood that when 2 individuals looked at the same sales data, one might have cared about whether the revenue hit a certain goal, while the other may have been looking for the most popular item. We quickly identify a goal to to provide personalized answers based on the things users cared about.

To understand whether users found these insights useful for their purposes, I created a way to allow users to provide feedback into the ML model we used. Drawing inspiration from highly personalized apps such as the Facebook Feed, Facebook Ads, and Instagram Ads, I developed 3 different options within Insights to provide feedback.

We conducted A/B testing for all three options and ultimately decided on option 3 due to its improved visibility to the users. This choice resulted in us receiving 10 times more feedback compared to the other two options. During the initial phase of developing ML personalization, this accelerated the training process of the ML model. However, given more time, I would like to incorporate the follow-up questions from option 1 into the design of option 3.

Excel Ideas successfully shipped to all Office users worldwide and hit 1 million MAU very shortly after the official launch. It was an amazing journey to work on something that reached so many users. Looking back at the user journey from the early research, it was evident that users always start from "asking a question". This inspired me to initiate a new project to help users proactively search for answers, rather than Excel attempting to guess their intention. I designed an experience inside Ideas that allows users to use natural language to ask Excel questions. Ideas finds the answer to the question and explains how it was calculated via a formula, a chart or a table. This not only helps users get answers that they care about quickly, but also helps them learn about creating formula, charts, etc. This feature is now also shipped to the public.

Through the years I got to know many other designers and product managers who worked in similar spaces for other products in Microsoft. PowerPoint shipped a "Designer" feature, which automatically turns boring slides into something that looks professionally designed. Word shipped "Editor", helping users not only on grammar, but also suggesting tones and writing styles based on whether users are writing an essay, formal letter, or a diary. Outlook, PowerBI, etc. all shipped similar features that help users to work more efficiently using AI. I was inspired by how everyone focused on the user needs, not what AI is capable of. However, I quickly realized that we needed our design to collectively present our brand and the identity of Office and Microsoft.

Gathering a group of talented designers owning various these products products across the company, we met weekly to share and discuss design patterns, visual, information architecture, and much more. We discussed topics such as "How might we enable users to provide feedback to suggestions", etc. Together we built the same card visual framework, navigation model, feedback mechanism, and visual elements such as color, typography, theme, etc. for a cohesive experience in Office. As part of this effort, we together defined the interaction behavior and design principles for different surfaces in the Office apps for these new intelligence capabilities.

As we continue to learn and evolve, surfacing suggestions on the side pane isn’t always the best experience especially when user’s in the middle of their workflow on the canvas, and our suggestion relies on the content user’s currently working on. I had an amazing team of PMs and engineers who were always supportive of me reaching out to users and explore new ideas based on feedback. Here are some of early concepts across different Excel surfaces, to help users become more productive with the power of artificial intelligence.