Two years of AI tooling has made one thing dramatically cheaper: producing screens. A competent designer can now put forty variations of a dashboard in front of a stakeholder in an afternoon. What has not changed at all is the cost of being wrong about which dashboard to build.
Cheap output raises the value of judgement
When variations were expensive, the constraint did some thinking for you. Producing three options forced you to decide what mattered first. Producing forty removes that discipline and quietly moves the bottleneck downstream, to the person who has to choose. Teams that shipped faster with AI are the ones who got stricter about the choosing, not the generating.
Where it genuinely helps
Research synthesis is the real win. Thirty hours of interview transcripts used to mean a week of tagging before anyone could see a pattern. That is now an afternoon, which means we run more interviews rather than fewer. The tooling did not replace the research. It removed the excuse for skipping it.
It is also good at the first eighty percent of a component: sensible states, reasonable spacing, a working keyboard path. That frees senior time for the twenty percent where a product actually differentiates itself.
Where it quietly costs you
Generated interfaces trend toward the average of everything the model has seen. That is exactly right for a settings page and exactly wrong for the screen your business runs on. We have reviewed products where every flow was defensible and none was memorable, and the team could not work out why nothing landed.
A model can tell you what a checkout usually looks like. It cannot tell you why yours loses people at step three.
Our rule is simple. AI touches anything commodity, and never the two or three screens that carry the product. Those still get argued over by people who have watched real users fail at them.