The interface 50 million packages a day depend on

Amazon's fulfillment centers process roughly 50 million packages every day. The associates doing that work spend their entire shift staring at a single piece of software: PackApp. It tells them what to scan, what box to use, what labels to apply, and when a shipment is complete.

PackApp hadn't been redesigned since 2013. In a decade, the workforce had diversified dramatically. Associates spoke dozens of languages. They had varying levels of tech literacy. Many had physical disabilities. The app had none of that in mind.

The numbers told the story. Pack Singles variable cost had climbed from $0.26 to $0.38 per unit since 2017. New associates needed 88 hours to reach veteran proficiency in Pack compared to 43 hours in Pick. Recordable injury rates in Pack were the highest of any process path in the network. The app wasn't just outdated. It was causing measurable business damage.

Before designing anything, I needed to understand what was actually broken. I co-led a 12-week discovery phase: visits to 12 fulfillment centers across North America and Europe, heuristic evaluations of three core pack process paths, stakeholder interviews across 6 Amazon teams, and usability sessions with 104 associates across 6 countries.

I conducted a full UI audit, mapping every screen across 20+ pack modes, cataloging components, identifying redundancies, and building an Associate Map: a complete view of every digital and physical step a packer takes in a single shift.

What I found wasn't a visual problem. It was a cognitive load problem at industrial scale.

PackApp had accumulated over 20 distinct pack modes over a decade. Each was a slightly different version of the same core workflow, built to accommodate variations in station type, region, fulfillment type, and process path. Every time Amazon added a process requirement, someone built a new mode instead of evolving the existing ones.

Associates switching between modes were consistently the ones making errors. The inconsistency was the defect source, not a side effect. Leadership disagreed. Changing software used by 3 million associates felt catastrophic if something broke.

That data shifted the conversation. Leadership approved consolidation. We moved forward with a single adaptive interface across all pack workflows, with context-specific variations handled in the logic layer, not the UI layer.

The design wasn't a single breakthrough. It was five compounding decisions, each addressing a specific failure mode identified in research. Each decision had a legitimate counter-argument. Each required evidence to hold.

Every decision I made affected millions of people doing physical work in high-stakes environments. That changes how you think about what good design means. It is not about elegance. It is about what happens at the tail end of the distribution: the associate who is 5 feet tall, packing for 8 hours in a second language, on their third station of the day. That person is your user. Design for them and you have designed for everyone.

The other thing I learned: seniority in design is not about taste. It is about knowing when to make the business argument, how to build the evidence for it, and how to hold a position when leadership says the risk is too high. Sometimes the job is to show that the real risk is staying still.