Copilot
your everyday AI companion
By the end of 2022, I felt like I was hitting a ceiling at Microsoft and needed a change to accelerate my learning. Then, out of nowhere, came ChatGPT—and with it, Microsoft Copilot. This was my dream: working on emergent technology.
It hasn’t been without challenges—one of the biggest being the need to connect and bridge teams across different products at Microsoft, as well as multiple teams within Microsoft Teams, to align efforts and create cohesive experiences. This work required close collaboration with diverse disciplines, including designers, engineers, PMs, and data scientists. The fast pace, coupled with confidentiality, often limited research opportunities. I’ve made mistakes and faced moments of frustration, but it has all been worth it.
As the scope of our work expanded, I helped onboard and mentor another designer who was focused on different Copilot projects. Collaborating closely and aligning on shared goals became an equally rewarding part of the journey.
Below, you’ll find a collection of projects in the Copilot AI space, ranging from helping users catch up on conversations to enabling clearer, more effective communication. Here’s a front-seat glimpse into Microsoft’s next big, bold step. Want the full story? Let's chat!
Catch up on conversations 2023
Products like Teams have made communication easier—both a good and a bad thing. The ease of communication has created the challenge of staying on top of numerous conversations. We recognized that AI could help lessen the burden by summarizing missed conversations. What began as a feature for providing conversation summaries evolved into a full-fledged conversational interface, allowing users to ask Copilot about anything within the conversation.
Copilot introduces a dedicated pane in each conversation, allowing users to not only summarize but also dive deeper into the conversation’s contents. The decision to use a pane as the modality was a collaborative effort with teams across Microsoft, ensuring consistency across the suite.
While I was part of the larger cross-Microsoft effort defining Copilot components, I simultaneously designed these components for Teams-specific experiences, serving as the bridge between Teams and Microsoft. The challenge was balancing what was best for Teams with the need for a consistent model across the suite.
What made this even more complex was that Teams was the only product in the suite with established patterns for chat experiences. Designing a ‘chat’ with Copilot meant walking a tightrope between adhering to these existing patterns and rethinking them entirely for the AI era.
Users could either type their own prompt or select from a set of pre-canned options to guide Copilot’s response. Every point in Copilot’s reply was accompanied by a citation, which not only referenced the source message within the conversation but also allowed users to navigate directly to it.
I worked on both the web and mobile experiences for this project. After taking it to public preview, I handed it over to a teammate to focus on launching Compose Copilot (helping users write and rewrite messages). Post-launch, I returned to it, exploring ways to drive engagement and retention, which led to initiating a new project, Proactive conversation summaries, presented later on this page.
Your writing assistant 2024
With over 2 billion messages sent per day, I was thrilled to leverage AI to assist users with their writing needs. Designing for mobile was particularly challenging as it required incorporating all functionalities into a significantly smaller footprint while maintaining usability and clarity. Compose Copilot is made up of three key features, launched in this order:
Rewrite: Rewrites messages you’ve written based on pre-set prompts.
Custom Adjustment: Rewrites messages based on a custom prompt.
Write: Drafts messages from scratch, based on a user-provided prompt.
Rewrite allows users to choose between a generic rewrite or one of the pre-set prompts for length or tone. The experience also lets users navigate between multiple outputs, enabling them to tweak the text to their liking and select their preferred version—a feature that was particularly appreciated by early users.
Early feedback from users revealed a desire to specify how they wanted Copilot to rewrite messages. This insight motivated us to accelerate the release of Custom Adjustments.
The ability for Copilot to draft messages from scratch felt like the feature that would truly complete the experience.
I thoroughly enjoyed designing the Copilot compose experience, but early on, I anticipated that it might not achieve the usage levels needed to justify further investment. The challenge was: Chat in Teams saw significantly higher usage than Channels, and messages in Chat were typically short, with the P90 being under 157 characters. This made it less productive for users to rely on Copilot to craft such brief messages, potentially slowing them down.
Early user feedback echoed this concern. We brainstormed ideas to address it, such as enabling Copilot to leverage context from the conversation to craft responses with minimal input and to learn and write in the individual user’s tone. While I enjoyed designing for these scenarios, we couldn’t convince leadership of the ROI to move forward.
That said, the feature has developed a small set of loyal users, many of whom are non-primary English speakers who love using it to improve their communication.
A collaborative agent for your team 2024
2024 began with a directive from leadership to deliver a collaborative AI agent—a vision that was undefined, with no clear problem to solve or understanding of how it would work. It was up to us to figure that out, marking the start of the most challenging project I’ve worked on. With intense pressure to deliver at breakneck speed, I led the collaboration across multiple teams. The journey was fraught with hurdles: we repeatedly hit a wall in defining the problem, making it feel like technology searching for a purpose. The technology itself wasn’t yet capable of fully realizing our vision, and while research provided insights into what didn’t work, it offered little clarity on what would.
We were attempting to challenge established patterns without a strong foundation of conviction. Despite the obstacles, we shipped a collaborative agent that could be added to conversations to assist groups with tasks. Its initial capabilities included summarizing content, finding files, and searching the web—just the beginning of agentic AI in Teams.
This project pushed me to grow in meaningful ways: learning to collaborate more effectively, stay calm under pressure, and embrace teamwork at a deeper level.
The collaborative agent is designed to actively participate in conversations, taking over mundane yet necessary tasks to let humans focus on what truly matters. For instance, it can summarize discussions, answer questions by pulling information from the web, and even analyze and respond to questions based on documents shared within the conversation.
Take the example of scheduling a meeting. Copilot could handle this task across reactive, reflective, and proactive scenarios:
Reactive: Someone in the chat explicitly asks, “Can we schedule a follow-up meeting?” Copilot responds by presenting available times for all participants, pulling relevant agenda points from shared documents, and generating a meeting invitation.
Reflective: Copilot observes ongoing back-and-forth about a complex topic in the chat. Recognizing that a focused discussion could be more effective, it suggests scheduling a dedicated meeting and offers time options based on participants’ availability and the urgency inferred from the conversation.
Proactive: Without any explicit prompt, Copilot identifies a high-priority topic spanning multiple shared documents and unresolved queries in the conversation. It proactively proposes a meeting, creates an agenda by summarizing key points from shared files, and suggests next steps to ensure alignment.
The introduction of Copilot as a shared collaborative agent for the team sparked a major challenge. It created confusion between the private Copilot users were familiar with and this new shared version, leading to extensive back-and-forth debates, multiple workshops, and involvement from leadership. While we haven’t yet landed on a fully convincing solution, these efforts moved us closer to the concept of One Copilot—working toward helping users understand the scope of each Copilot within its specific context.
Another significant discussion centered on whether AI agents should be treated on par with humans in conversations—both in terms of the mental model and the resulting UI patterns. Research showed that people didn’t yet perceive AI agents as equal to humans (understandably). However, we chose to take the first step toward shaping this future by placing these agents in the same menu as the participant roster.
We weren’t just designing and building for a single AI agent but for an entire platform to support a variety of agents. It was my responsibility to define scalable patterns that could be integrated into this platform, enabling seamless support for 2P and 3P partners.
Proactive conversation summaries 2024
I noticed a recurring pattern: we’d ship a feature, it wouldn’t see the expected usage, and PMs would push for better discoverability—often through prominent entry points or coach marks. But I was convinced that the reason people weren’t using Copilot in chat wasn’t due to lack of visibility. Instead, they lacked a mental model of integrating AI into their workflow for catching up on conversations. Surface reminders might help once or twice, but we needed to go further.
The solution was to bring intelligence to where users are by proactively surfacing the most helpful, scenario-specific content without requiring them to seek it out on an AI-dedicated surface. I identified two key scenarios: when users entered a conversation with numerous unread messages and when they joined a conversation they’d just been added to. This sparked an exploration of how to effectively surface these AI-generated summaries in a way that felt intuitive and valuable.
We experimented with several ways to present the information, ultimately opting for a prominent placement above the compose bar in an all-or-nothing approach. The goal was to understand whether users would find it helpful or if it would disrupt their workflow. We had a preview audience to test these ideas and gather insights. While I was eager to see this through and learn how users reacted, I left Microsoft before I could take it to completion.
Channel announcements 2024
Data showed that channel posts, particularly announcements, took users the longest time to draft. These posts were typically longer and often included a banner image. This insight sparked the idea that using AI to generate a first draft could significantly speed up the process. It would be a game-changer if Copilot could leverage its knowledge from project conversations to create a contextually informed draft. While this wasn’t an official project, it was something I worked on as a side initiative.
✌
Thoughts from the frontlines
Working on Copilot was a transformative experience. It pushed me to navigate complexity, align diverse teams, and guide collaboration across Microsoft, shaping how I contribute and support others in challenging environments. The fast pace of innovation gave me greater autonomy and taught me to make quicker, better decisions—though not without mistakes, which became valuable learning moments.
The project also sparked a desire to work at a smaller company with less bureaucracy. While we moved swiftly to keep up with technology and competitors, the layers of decision-making in a large organization often slowed us down. I’ve come to value agility—the ability to focus on building rather than managing processes.
When I first started in the AI space, it felt unproven and disconnected from real-world needs. The technology struggled to meet aspirations, and progress felt slow. But I’ve witnessed its rapid evolution, far exceeding expectations. What amazed me even more was how eager people were to adopt this new technology, defying the usual resistance to change. Early data revealed not just its potential but also the readiness of users to embrace it, which fundamentally shifted my perspective on AI’s impact.
That said, much work remains. Many experiences still feel unfinished, and we need to ensure the solutions we build genuinely solve user problems, rather than feeling like technology in search of a purpose.