The part that got fast
For most of my career, the bottleneck in design was production. Ideas were plentiful and screens were expensive, so the person who could turn one into the other held the scarce skill. That bottleneck is gone. A machine will now produce a plausible flow, in my style, with sensible copy, in the time it takes to describe it. If your value was the plausible flow, this is a bad decade to be you.
I keep coming back to the sentence this whole site is built on: interfaces are where decisions become visible. The machines have made the visible part nearly free. The decisions did not get cheaper. Knowing what to build, for whom, and what it should teach people to expect — that work moved upstream, and it is now most of the job instead of the part before the job.
How my week actually changed
The prototype comes before the meeting now, not after it. I used to spend the meeting arguing for the idea and the week after building it; now I build three versions before lunch and let the best one make the argument. I have always believed the best prototype changes someone’s mind. The change is that a mind-changing prototype is now hours away, which means alignment, the slowest thing in any organization, got a faster tool.
Writing became the load-bearing skill. Prompts are writing. Specs are writing. Evals, the checklists that define what good means so a machine can be graded against them, are writing with consequences. The designers I see thriving are the ones who can say precisely what they want and precisely why a result isn’t it. Critique, in other words. The editing room matters more than the drafting table.
And teams are getting smaller and stranger. When one person with good judgment can move at the speed a pod used to, the interesting question stops being “how do we staff this?” and becomes “whose judgment do we trust with this?” I find that clarifying. Judgment was always the product. Now it’s legible.
Designing things that guess
The other half of this is designing for AI, and here I have scar tissue. At Airbus Aerial we shipped machine assessments to insurance adjusters years before language models made “confidence” a design word. What I learned there still holds: a system’s claim is not an answer, it’s a claim. It needs evidence, provenance, and a cheap way for a human to disagree. Percentages invite false precision. Trust is calibrated, not declared.
A product that guesses should show its work, admit uncertainty gracefully, and make disagreement cheap.
Every product teaches people how to behave, and AI products teach people what to trust. That’s a heavier syllabus. A system that flatters you teaches you nothing. A system that’s confidently wrong teaches you the wrong thing at scale. The hard design problems of the next decade are not layout problems; they are epistemics problems wearing a layout problem’s clothes. This is exactly the territory I want to work in, because it’s where finance was when I got there: powerful, opaque, and badly in need of someone asking what the person on the other end actually understands.
What doesn’t move
People are still confused by their money at 2am. Executives still decide in hallways. Organizations still ship their org chart. None of that changed because the tools did, and none of it is solved by generating screens faster. The work underneath design — understanding people, finding hidden assumptions, creating alignment — got more valuable precisely because everything above it got automated.
So I’m not nostalgic and I’m not breathless. The tools changed. The question didn’t: what are we actually trying to build, and will the person on the other end understand it? Machines are getting very good at answers. Someone still has to be good at the question.
Disagree with any of this? Even better. Schedule a conversation.