What AI Actually Changes About Web Design

There's a version of this post we could have written eighteen months ago, and it would have been wrong. It would have said AI was going to either replace web designers or change nothing, and the truth turned out to be neither. After a year of using these tools on real client work — not demos, not Twitter threads, actual sites we shipped and stand behind — here's what actually changed and what didn't.

What changed: the cost of the first draft

The biggest shift is that the first version of almost anything is now nearly free. A first pass at copy, a layout direction, a set of naming options for a content model, twenty variations of a hero headline — these used to cost an afternoon. Now they cost a minute.

That sounds like a productivity story, and partly it is. But the more interesting effect is on how we work. When the first draft is cheap, you stop being precious about it. You generate three directions instead of agonizing over one, you throw two away, and you spend your saved time on the part that was always the hard part: deciding which direction is right and why.

The discovery and brief still matter more than ever — arguably more, because the tools amplify whatever you point them at. Point a model at a vague brief and you get a confident, polished, vague result. Garbage in, fluent garbage out, which is more dangerous than the old kind because it reads well.

What didn't change: the work that was always the work

Here's what AI has not done for us, not once: it has not decided what a client's business actually needs. It has not sat in a founder interview and noticed the thing the founder didn't say. It has not looked at a content audit and understood that the real problem isn't the copy, it's that three departments each think they own the homepage.

Judgment didn't get cheaper. If anything the gap widened between teams that have it and teams that don't, because the teams without it now ship their bad judgment faster and more confidently.

The same goes for the craft floor. A model will happily produce a layout that technically works and has no point of view. It will write accessible-looking markup that fails a real audit. It will suggest a component pattern that's fine in isolation and wrong for the system it lives in. Catching all of that still requires someone who knows what good looks like — and knowing what good looks like is not something you can prompt your way into.

How we actually use it

We're not purists about this. A few places it earns its keep on real projects:

First-draft copy and content modeling. Especially the unglamorous parts — field labels, microcopy, the forty product descriptions nobody wants to write. We edit heavily. We never ship it raw.
Code we already understand. It's a faster way to write the boilerplate we'd write anyway. We don't use it to write code we couldn't have written ourselves, because we can't review what we don't understand.
Interrogating our own thinking. “Argue the opposite of this design decision” is a genuinely useful prompt. It's a cheap second opinion, and one that never gets tired or political.

And a few places we keep it out:

Anything load-bearing we haven't verified. Accessibility, performance budgets, security — the model is a starting point, never the sign-off.
The voice work. When we build a voice guide so a client's baristas and marketers sound like the same company, that's the product. Outsourcing the soul of it defeats the point.

The honest bottom line

AI didn't replace web design. It replaced the cheapest, most mechanical 20% of it, and in doing so it made the other 80% — the judgment, the taste, the actually-understanding-the-client part — more valuable, not less. The studios that struggle with this won't be the ones that refused to adopt the tools. They'll be the ones who mistook a faster first draft for a finished one.

We bill for the judgment. We always did. The tools just made it more obvious that's what you were paying for.