Decisions that took an afternoon
Three decisions that would previously have taken a quarter of debate, settled quickly because we could put the real thing in front of people. I now run the team on decision latency.
Collecting three from the last few weeks, because they share a shape I want named.
Mobile-first signing. New hires sign documents during onboarding, and many do it on a phone, standing up, mid-shift. Previously this would have been a desktop design followed by a mobile adaptation, and the adaptation would have lost every argument it had. Instead we built both and held them. The phone-first option won on evidence, in one session, because everyone in the room could pick it up and try to sign something.
Base font size. Fourteen pixels or sixteen — exactly the kind of question that becomes a taste argument between confident people and consumes a month. We kept sixteen and added user-level text size control, which is a better answer than either number. We got there quickly because the disagreement was about something running rather than something described.
Drift flowing the right way. Someone synced their project against the template, the tooling flagged they had drifted, and they opened a pull request back into the template. I had built a one-way distribution mechanism and what we got was a loop. The old process had no cheap path from "I hit this while building" to "the shared thing is fixed". Now it is a pull request, and fixing one pattern today means every screen built after it starts better.
The common factor is not speed for its own sake. It is that each of these became answerable the moment there was something real to look at. A question about a drawing gets settled by whoever is most senior or most persistent. The same question about a running thing gets settled by the thing.
So I have changed what I measure. Not throughput, not velocity, not how many screens a designer produced. Decision latency — how long between a question being raised and being answered with evidence. It is the number that predicts everything else, and it is the one I will use to run design organisations from now on.
Agents are part of this, in a less glamorous way than people expect. During generation they check contrast, labels, token use, light and dark. They are faster and more consistent at that than any person, including me, and I would not go back. What they cannot tell you is whether someone using a screen reader can finish applying for a job, recover from a confusing error, or find the document they need before a shift starts.
That is the future I am building the team for. As generation gets faster, the gap widens between what got built and what anyone actually verified — and the people who can close that gap become the most valuable people in the room. So I am not training designers to produce more screens. I am training them to answer harder questions faster, and letting the machines do the part that was never the point.