Driving global impact: leading mixed-methods research and AI innovation at Apple.

The most important finding was about confidence. People could complete their tasks, but the system never confirmed that a correction had gone through. When the task is fixing your own pay, that uncertainty stays with you long after you close the screen.

I turned the findings into a ranked list, organized by root cause. Internally, it shaped the training curriculum, set the communications timeline, and got confusing terminology renamed before launch. With the vendor, I let the data make the argument: how often each issue occurred, how often people failed the task, and where the cost would land. A task people can't find, for example, becomes ticket volume for the People team. Every item was a specific, scoped fix instead of a redesign request, and several shipped, including some that had been ruled out of scope.

After launch, a new scheduling screen shipped with five color-and-pattern combinations displayed at once. I tested it with five people who use that screen all day. Within seconds, all five reacted without any prompting, and three described physical symptoms. The screen passed WCAG, which measures whether colors contrast enough to tell apart, but WCAG doesn't account for chromatic vibration, the shimmering effect created when strong patterns sit side by side. I grounded the finding in published vision research by Ware and Wilkins and in the ISO 9241-112 standard, so the case rested on evidence instead of taste.

The vendor accepted the finding and adopted the term. They named pattern reduction as the leading fix and put the release under review while they worked out what to change. Not every change has shipped yet, but a roadmap that had been treated as closed was reopened by a customer with no authority over the product, on the strength of the research.

The same pattern held across all three moments. Multi-token search built on the mental model people already had, and the response was immediate: "It is amazing to add multiple tokens. Ship it. I want it." Later, a change to core search behavior shipped without research and contradicted that model. The Voice of Customer program measured the cost: 372 responses across 24 countries, 57% satisfaction, and 28 comments describing it as a regression. The AI chat study confirmed the pattern a third time. All 11 participants reached for short keywords, welcoming the AI where it made something new possible and resisting it where it added a slower path to a familiar task.

Eleven participants can explain a behavior, but they can't move an AI roadmap on their own. So I brought the study to AI leadership alongside outside market data: Gartner's projection that search volume would drop 25% by 2026, and a March 2026 survey showing early enthusiasm for AI-first experiences settling into more selective use. The industry pattern mirrored my sessions closely. People preferred AI for step-by-step tasks and chose search 76% of the time for familiar queries.

My recommendation was to build the AI and launch it as an opt-in feature. Keyword search would remain the primary entry point, with a single input that accepts both keywords and natural language. That approach gives experienced users a gradual transition and lets the feature earn its adoption over time. It also shaped the principle I now lead with: adoption depends on whether a design extends or contradicts people's existing mental model, far more than on how impressive the technology is.

I conducted this study using the AI research skills I built for our team, which reduced synthesis and reporting time by 75% across our work. Using AI to evaluate an AI feature proved valuable in its own right. It means I can speak to where AI genuinely strengthens a research workflow and where it needs guardrails, from the experience of building those guardrails myself.

Time turned out to be the constraint behind every choice. Employees have a short window to get food and eat it, so they choose whatever is fastest and closest. Kiosk users believed app orders were prepared faster, and some described arriving late to lunch with colleagues whose food was already ready. The kiosk was often the slower option, chosen anyway because it was the predictable one. That made the highest-value recommendation a simple one: show estimated wait times at the kiosk so people can make an informed choice.

The finding that shifted the most thinking was one the study wasn't designed to find. Four of fifteen participants didn't own an iPhone. Stakeholders had treated the app as universally available and kiosk use as a preference to correct, but for a meaningful share of employees, the kiosk is the only way to order. That single data point reframed kiosk investment from a legacy cost into an access requirement, and it surfaced only because recruiting wasn't screened by device.

I defined three user types to give the redesign a clear target: App-only, Hybrid, and Kiosk-only. Kiosk-only users optimize for the fewest possible taps, while Hybrid users arrive expecting app features like Buy Again and Order Ahead. I also documented that kiosk users browse the menu on the web or ask friends before they arrive, which makes the web menu an essential part of a journey that hadn't been part of the design work.

For team members, simply staying on top of their own schedules took real effort. They checked the time app 6 to 8 times a day, kept handwritten whiteboards, and set personal alarms to catch their own breaks, because the tools sent no notifications. The baseline survey showed where that friction concentrated: 53% disagreed that schedule edits are easy, and 43% of 1,130 comments focused on editing difficulty. Satisfaction sat at 49% while confidence reached 63%. People knew how to use the tool and still found it hard, which pointed to specific features that needed fixing, beyond anything training could solve.

Some of the most consequential findings fell outside the original brief. Stores were informally scheduling around language skills and accessibility needs, with no support from the system: "The store must have 1 Spanish speaker at all times on-point, but it is never on schedule." Two Deaf team members were repeatedly scheduled as phone operators. Those findings reframed scheduling from a convenience tool into an equity and compliance issue.

On device management, the recurring theme was that official guidance wasn't always feasible in practice, so teams built workarounds to close the gap, including imaging devices in unsecured areas because the secure room lacked the network capacity to finish in time. I placed that directly in the findings, alongside launch timelines that assume nothing goes wrong and guidance that overlooks regional limits on overnight work. Framing the findings as opportunity areas made them something a 2 to 5 year roadmap could actually build on.