Google study: expert-built UI still beat the machine, even when the model got faster

The work you own is moving off the screen. Your team still ships buttons and forms, but the harder decisions now sit in what the software infers, what it does on its own, and where a person steps back in. Three signals landed this month that say the design job is being redrawn. Let me catch you up.

The deep cut

• Design owns the decision, not the pixel. Google's own study found expert-built UI still beat generated UI, even as the model got faster.
• Put the human before the cost, not before the ship. Jeff Gothelf's Claude agents ran wild for a month because the checkpoint sat too late.
• A neutral parts list beats a rented one. Slack's Block Kit and Microsoft's Adaptive Cards both draw only inside the owner's walls.

Same form, four times over

Your design system already serves web, iOS, and Android. Now add ChatGPT, Claude, Gemini, Copilot across Teams and Outlook, and Slack. That is the same date picker built over and over, once in Block Kit JSON, again as an Adaptive Card, again for whatever ChatGPT wants. Ten surfaces times three platforms is not ten problems. It is a matrix nobody staffed.

Every assistant draws its own way, and the bill lands on your design systems team. Slack controls contrast so you do not have to, which is a gift right up until the client belongs to someone else. If you have ever fought a Block Kit layout into your brand, you know the ceiling of a parts bin somebody else owns.

Google's A2UI is the one proposal on the table that no vendor owns. The agent names components the client already holds, and styling stays with your system. The catch is honest: the mobile renderers are mostly volunteer work, and the spec is still a release candidate. Do not bet the roadmap on it yet. Do start tracking it.

The screen stops telling the whole story

Fixed software waits for clicks. Agentic software reads a goal, plans steps, and comes back with a result. That changes what the interface has to carry. Lucas Camara's ERP example makes it plain: instead of picking a unit, a date range, and five filters, a manager asks why the result fell versus last month. The system does the assembly.

That does not delete the ERP's complexity. It moves it into interpretation, where mistakes are harder to spot. So the interface has a new job: show which records were included, which definition of "result" it used, what data was missing, and how the math was done. When the system acts, permission, confirmation, and a reversible record become required, not nice-to-have.

Treat trust as parts you design, not a feeling you hope for: permission, provenance, confirmation, receipt, reversibility, escalation. And build for uneven ground. Gartner expects task-specific agents in 40% of enterprise apps by the end of 2026, and also expects more than 40% of agentic projects to get canceled by the end of 2027. Cost and weak risk controls kill them.

The human who only clicks approve

Here is the failure mode nobody plans for. Jeff Gothelf took August off and came back to a shitshow of broken Claude tasks, including an agent that built a parallel system to duplicate his own because it could not get answers. A human in the loop only helps if the human is in the right part of the loop.

Most checkpoints fail for three reasons. The reviewer gets the output, not the reasoning. The check comes at the end, when starting over costs a week, so approve is the only real choice. And there are too many reviews per day, so people learn to skim. That is a rubber stamp wearing a lanyard.

Gothelf's fix is a rule you can repeat: put the human at the last cheap moment, one step before reversing course gets expensive. In his content workflow, that is right after the agent narrows to one idea, not later when the draft is written. Ask where course correction gets costly, and plant your reviewer just before it.

What the designer is actually for now

The visual job is shrinking and the judgment job is growing. The U.S. Bureau of Labor Statistics projects graphic design jobs to fall 2% by 2035 while web and interface design grows 6%, and the pay gap says the same thing: $62,960 median for graphic designers, $104,000 for digital interface designers. Figma's hiring research found 79% of managers want knowledge of AI product design.

The proof this is judgment work, not pixel work: Google's November 2025 evaluation found people preferred generated UI to plain model answers, but expert-built interfaces still scored highest, and generation could take a minute or more. A machine can make the artifact. It cannot decide which one is worth shipping.

Flexy Global put the enterprise version in practice: make every recommendation traceable, and lower the agent's autonomy as the cost of a wrong answer rises. Their test for whether trust is real is three steps, Understand, Question, Act. A finished task does not prove the person understood the answer.

Three questions for your team

• Which agent surfaces do we commit to, and which do we refuse, before our design system quietly signs up to serve all ten?
• In our top AI workflow, where does reversing course get expensive, and is our human checkpoint sitting one step before that or at the very end?
• When our system acts on a user's behalf, can they see what it will do, stop it, and undo it, and can we show that in a review next week?