William Lee

AI Product Designer

I am an AI product designer. I design the interfaces that make machine intelligence trustworthy, and I run the experiments that prove whether they work. William Lexen is my product studio: the studio name is on the door here, my name is on the resume, and this sentence is the bridge between them so a person arriving from one does not think they have found a stranger.

I have designed product for fourteen years, and eight of those went into a single real estate analytics lineage called BrokerMetrics, which I carried through an acquisition when Terradatum was bought by Lone Wolf Technologies. I am still its design lead. That is an unfashionably long time to spend on one product and it is the most useful thing about me. I took a twenty year old platform and led its modernization into a scalable web product, which meant owning the onboarding, account creation, and migration flows that had to carry two decades of existing customers across without losing them, and it meant building the design system that holds the dense data and the workflows together. What happened there is confidential, so I will describe the work and not claim a number for it.

What eight years on one product teaches you is narrow and it is worth more than range. Dense data in front of a person who has to make a call under time pressure is a specific problem, and you do not learn what it does to someone by visiting it for a quarter. You learn it by watching the same broker, on the same dashboard, defer the same decision for the fourth year running.

Because here is what I kept seeing. Most software shows you everything and tells you nothing. It hands you numbers, charts, and trends, then goes quiet at the exact moment you need it to tell you your next move. You do the math in your head, you defer, or you hand the decision to whoever sounds most confident in the room. I watched that happen for fourteen years on products where the stakes were real, and somewhere in there I stopped believing it was a data problem. The information was never missing. It was unreadable.

So I wrote it down. Death by a Thousand Decisions is a field method for designing screens that help a person make the call they came to make. It gives you CRAFT, which is five questions every high-stakes screen has to answer before it earns the right to show you data, and it gives you Interface IQ, which is a way to score any screen from 0 to 100, name exactly what is dragging it down, and prove it got better after you fix it. The point of a number is not the number. The point is to stop losing design arguments to whoever is most senior and start settling them with a diagnosis and a fix. Then I turned the instrument on two production dashboards and measured before and after.

Those two pairs are the only fielded numbers on this site, and they are here because they have data behind them. Everything else you might want a number for, I do not have one for yet, and I would rather say so.

Which brings me to the part that is actually in flight. A book that argues you should score an interface instead of arguing about it is only worth something if its central claim survives contact with reality, so I am running two pre-registered experiments against it, on the surface where I think the argument now matters most: interfaces that agents have to read, not just people. That is what AI product design is, as far as I can tell. Not a coat of chrome on a chatbot, but the question of whether a person, or a model acting for them, can tell what a system knows and how much of it to believe.

A five star rating cannot tell you whether the shipping was fast or the product was good. It is equally useless to an AI shopping agent, which cannot ask a follow up question. Agentic commerce is arriving and the review infrastructure of the entire internet is unreadable to it. That is the first study. The second asks whether agent-readable form design is a real discipline or just good form design in a costume, which is the version of the question that could embarrass me, and therefore the only version worth running.

Neither has run. The hypotheses, the methods, and the falsification conditions are committed before the data exists, which is the whole point and the reason there are no results on this site yet. Most designers hand you an opinion. I would rather hand you a number and the pre-registration that proves I did not go looking for it. When the results land they go on artifacts, including the nulls.

I can run them because I build them. I design in Figma and validate in code, and I ship production software solo: React, TypeScript, Next.js, Tailwind, Cloudflare Workers. That is not a detour from design, it is what lets me find out whether a decision was right instead of arguing about it. Most senior designers hand engineering a file. I hand them something that already runs. It is also how the studio works, which since 2025 has shipped native iOS, booking platforms, and data tools end to end, and how I work now with Bosko Partners on Saurara, standardizing a multi-role survey and analytics platform whose admin, organization, and survey surfaces had drifted apart from each other.

None of this is a late conversion. I was a creative director at VScreen through 2018, running UX strategy for SaaS video products used by real estate agents, and I was using behavioral analytics to decide what to change about a decade before that was fashionable. Before that I was a front-end developer. The measurement habit came first, the design system habit came second, and the book was just the first time I wrote the argument down where someone could check it.