Search with Copilot
Transforming Outlook search with AI
Lead designer
2023 - Ongoing
How might we help users efficiently find what they are looking for by infusing AI into the search experience?
1. HOW IT ALL STARTED
We began this work in a space of pure exploration. We had this incredible new tool - AI - and we had no idea where the constraints lay.
Our mandate was simple: ignore feasibility and cost and focus on delivering an experience that felt like magic to the users.
I immersed myself in rapid design sessions, imagining every possible usage of AI in the search flow.

This stage was extremely important for anchoring the project in user value, not technical limitations. We needed to know what users wanted before we asked what engineers could build. We took these high-level, speculative designs directly to users to gauge their genuine reaction—desirability was the only metric that mattered at this point.
Early drafts - 2023
2. KEY CHALLENGE
Designing in the Dark
We kicked off this work in early 2023, immediately following the public debut of ChatGPT.
At that moment, Large Language Models (LLMs) were far from mainstream, and very few of our users had any practical familiarity with conversational AI concepts.
Because users were not familiar at all with anything we were showing, they were often confused by the new AI features and frequently expressed unhappiness with every direction we presented.
We knew any major UI change invites the predictable "Don't move my cheese" resistance, and anticipated that introducing AI would amplify this pushback.
For this reason, we decided to take initial negative feedback with a necessary grain of salt. Our core hypothesis was that repeated exposure would lead users to embrace the new AI capabilities and ultimately adopt the new search paradigm.
Instead of halting momentum due to early resistance, we focused on maintaining forward motion with high-conviction ideas, trusting that long-term utility would overcome short-term discomfort.
3. STRATEGY
Grounding the vision
Once we had a pool of high-value concepts, the story shifted. We had to move from 'What if?' to 'How do we start?'
To ensure that Design consistently led the way while simultaneously guaranteeing that Engineering could immediately begin building what was technically feasible, I had to shift my focus and lead three parallel workstreams at once:
Crawl
Designing what was feasible right now, allowing PMs to measure impact and engineers to ship quickly.
Walk
The incremental adjustments needed for the existing search interface to eventually support the North Star (Run) design.
Run
Defining our ultimate, unconstrained vision for AI-powered search.
This approach allowed us to maintain aggressive momentum, address present-day constraints, and keep our eyes fixed on the future vision, ensuring a phased yet deliberate approach to product evolution.
CRAWL
To get the AI capability into users' hands quickly and gather real-world data, we defined a simple MVP: a single, clear button placed prominently above the traditional search results. This button allowed users to trigger a Copilot search based on their original query.
Windows MVP. Shipped to WW in 2025
The primary goals of this intentionally minimal launch were:
Measure genuine user interest and adoption.
Observe and mitigate model risks like hallucinations.
Work through known service issues such as latency.
Ensure high prompt relevance
Simultaneously with the Windows launch, we ensured this new experience achieved maximum user reach by shipping the same core UX across all major platforms: Mac, iOS, and Android.
WALK
For the next planned stage , we are preparing to introduce new concepts and greater complexity, beginning with a major overhaul of the search Information Architecture.
To accommodate the dedicated Copilot pane and ensure UI consistency across Outlook, we are repositioning core search features.
This structural shift is designed to support our long-term goal: encouraging users to abandon the single-keyword pattern that has dominated search for years.
By actively prompting and guiding users toward using natural language in their queries, we anticipate generating significantly richer input. This, in turn, will allow Copilot to deliver far more accurate and highly relevant results.
Ou goals:
Repositioning of tabs and refiners, to structurally accommodate Copilot wide pane
Implement AI-powered suggestions into the search bar to guide users toward using natural language in their queries.
Leverage richer, natural language queries to enable Copilot to deliver more accurate results.
RUN
Our ultimate North Star vision remains a work in progress. It is not a fixed destination but a living document that we continuously adapt. This adaptability is fueled by the insights we gain from our live experiments and user feedback.
We are currently engaged in a cycle of rapid iteration, constantly readjusting our designs as we learn from how users interact with the deployed AI capabilities.
This ensures that the vNext design we are building is not based on assumptions, but is directly informed by real-world behavior, keeping us responsive to evolving user needs and the true potential of the AI service.
4. CONCLUSION
What I've learned so far
Working with a new technology like Copilot has been an unparalleled challenge. Every time we introduce a new AI feature, we are doing more than just rolling out functionality; we are asking users to reimagine their workflow and fundamentally change search behaviors they have fostered for years.
This difficulty is amplified because we are dealing with a tech that is evolving incredibly fast but still has known flaws, such as hallucinations or latency. This creates friction, making it common for users to get frustrated and refuse to adapt. We are effectively building a plane while it is flying.
Consequently, the design process has proven incredibly demanding. We are simultaneously facing pushback and skepticism from users who resist change while dealing with significant restrictions and constraints from Engineering related to integrating complex new backends.
Despite these hurdles, I am genuinely excited to continue this journey, observing not only how the feature will evolve but also how user behavior and acceptance will evolve alongside it.
IT TAKES A VILLAGE
PM
Nima Dani
Priya Ganta
Engineering
Akashdeep Singh
Amelia Bateman
Anwesha Sarangi
Ivy Li
Joe Flint
Lawrence Yuan
Nathan Novielli
Ying Huang
Ying Zhong