Your screens are getting demoted. What are you defending?

You have spent years shipping screens. Flows in Figma, states, specs, happy paths, edge cases, handoff. The line we all repeated was that UX isn't really about screens. AI is now testing whether we meant it.

Agents are starting to sit on top of your product and do the clicking for the user. That changes the first question in every review, from "how good is this flow" to "should this flow exist at all." Let me catch you up on what's actually shifting, where your craft still holds, and what to bring to your next crit.

The steps nobody wanted to do

A lot of screens only exist because software couldn't act on what someone wanted. The config page with forty fields. The setup wizard. The dashboard with twenty metrics that's really a shrug: we're not sure what's wrong either, maybe you can tell us. Serg Zorin at Meta is blunt that much of his career went into screens built to fill that gap between intent and outcome.

That gap is closing. Microsoft took the internal form employees use to register a new device and rebuilt it as a chat agent, aiming to turn a six-step process into one or two actions. They kept the old form running, and usage is already dropping as people switch. Gartner expects around 40% of enterprise apps to ship task-specific agents this year, up from under 5% a year ago.

Complex tools feel this first. Power users never loved your dense interface, they tolerated it because the result was worth the clicks. Now they're asking if there's a skill or agent that does the job without the UI.

The screens that earn their keep

Most screens are getting demoted, not deleted. They move from where the work happens to a supporting role: the confirmation before something irreversible, the audit log, the report you check after, the fallback when the agent isn't sure. Still real design work, and some of it harder than before.

The screens that survive do judgment, oversight, and recovery. The dashboard you open during an outage to figure out what broke does a different job than the one that just lists metrics. Amelia Wattenberger's point holds here: an empty text box tells you nothing about what it can do, and bouncing between asking the AI and checking its work kills any flow. Thinking screens stay. Doing screens go.

There's a cost buried in the wins. When people stop performing the steps, they stop building a model of how the system works. That's fine until the agent gets stuck, and now the user has to recover from something they never learned. Counting removed clicks is a lazy scorecard.

When AI flattens the person

The hard part isn't generating a screen. It's holding complexity the model wants to smooth away. Johannes Seemann, building the finance tool Sooner, names three traps worth writing on your wall.

Normative contamination: a model trained on broad data projects value judgments onto what people say, and those get baked into a profile that shapes every recommendation. False coherence: models want to resolve contradictions into a tidy story. One user wanted to be his own man in New York while his parents covered part of his rent. Both true. A model under pressure to be coherent flattens one out, usually the nuance that mattered. Context rot: the longer a conversation runs, the more the model's early understanding fades, right when the decision gets bigger.

Seemann is also deliberate about control. When his team tested a free-flowing AI conversation against a push-to-talk format, people preferred push-to-talk because it handed control back. The AI surfaces options. The person decides.

Where the tools actually land

So how close is AI to real screens? Eldad pointed five AI UI generators at the same legacy banking screen, one prompt, first output, no cherry-picking. The gap between them wasn't styling. It was whether the tool reasoned about what a person needs from the screen or just reskinned the table.

Flowstep surfaced the next bill due date on the balance card, which wasn't in the prompt. v0 added Income and Spending summary cards, a real read on cash flow. Google Stitch grouped activity by Today and Yesterday. Lovable had good instincts but shipped a bottom nav that overflowed the viewport, the kind of thing a review catches and a first generation shouldn't. Figma Make kept the brand and hedged with a Raw toggle, reorganizing more than rethinking.

Which raises the Figma question. Nick Babich argues Config 2026 pointed at niche features like Motion and Shaders, disconnected from code, while the tools that turn a prompt into mergeable React are eating the handoff. If your design file is a dead end and the generator ships components, the center of your workflow is moving.

The deep cut

The skill that pays off now is the one that's hardest to automate: knowing when the machine is lying to you with a clean story. A model will flatten a contradiction, bury a value judgment in a profile, and forget what someone told it an hour ago, and it will do all of it while looking confident and polished. Your job is to design the moments where the person stays in control and can verify what happened. Push-to-talk over free flow. Confirmation before the irreversible. An audit trail when confidence is low.

So stop grading flows by click count. Start grading whether someone reached the right outcome with enough understanding to stand behind it. That's the question the generators can't answer for you, and it's the one worth defending your headcount on.

Three questions for your team

1. Pick your three densest screens. For each, can the system infer the answer instead of asking, and if the agent does it, where's the confirmation, undo, and audit trail? That's your list of what survives and what gets demoted.
2. When our AI features hit a contradiction or a low-confidence call, do we hand control back to the user or paper over it? Test a push-to-talk style handoff against a free-flowing one before you ship.
3. If a generator can turn a prompt into mergeable code, what part of our Figma-to-dev handoff still adds value, and where are we just moving pixels a tool now moves for free?