What AI changed about interface work

Interface design is the work of deciding how a piece of software behaves for the person using it. It covers the flow through a task, the screens along the way, the components they are built from, the words on every button, and the states nobody photographs, such as an empty inbox or a failed payment. Machines now touch almost all of it. They draw screens, write labels, turn a design file into code and sum up interview notes. This lesson looks at what that changed, what it left alone, and how the people who do this for a living have adjusted.

The first change is price. A high-fidelity screen used to cost a designer hours, so teams built habits around that cost. Wireframes existed partly because finished screens were expensive, and stakeholders learned to read a grey box as unfinished and a polished screen as decided. Today Figma's agent, Figma Make, Vercel's v0 and similar tools will produce a polished screen from a sentence in a few minutes. So polish has stopped being a signal. A screen that looks finished now tells you nothing about whether anybody has thought about it, and a client who sees one will assume somebody has.

What did not change is deciding what to build. Ask a model for a settings page, a checkout or a dashboard and you get the average of every settings page, checkout and dashboard it has seen. Sometimes that is exactly right. A settings page does not need an idea in it, and people move faster through patterns they already know. Sometimes the average is the problem, because the product exists to work differently, and the model has no way to know that unless you tell it. Deciding which case you are in is a product decision, and it is still a person's job.

The second change is where the value sits in a file. A screen generated with no design system behind it is a pile of loose rectangles with made-up spacing and colours. The same request made against a real system comes back built from your components, at your spacing and in your colours, because the system has already made the decisions the model would otherwise invent. That is why a clean library now outweighs any one screen drawn by hand. Figma publishes guidance on setting up a library so that its AI can understand it, and its MCP server lets a coding agent such as Claude Code or Cursor read a frame's components and variables directly instead of guessing from a screenshot. The file has become something a machine reads. Build it to be read.

The third change is the handoff. Designers used to pass a file to a developer and lose some quality in the translation. Now a coding agent can build a working version of a screen, or a clickable prototype wired to real data, in an afternoon, which is a real gain for testing ideas early. It also brings a new failure. Agent-built interfaces tend to look finished and be built wrong, with colours typed in as raw values instead of tokens, a new button where one already existed, and labels missing for people who use screen readers. None of that shows in a screenshot. So designers now read the result the way an editor reads a draft, at every width, with the keyboard and in dark mode.

Surveys show the same split between speed and trust. When Figma asked 2,500 of its users for a report published in April 2025, 78 per cent agreed that AI significantly enhanced the efficiency of their work, yet only 32 per cent said they could rely on its output. Developers were happier with their tools than designers were, and far more of them used AI in the core of the job. Its 2026 report, built on 8,403 survey responses, found that the share of designers taking part in development had doubled in a year to 41 per cent, while 90 per cent of respondents said design was at least as important as it was before AI. The drawing got faster. The deciding did not.