What AI changed about design, and what it didn't

A few years ago I sat in a meeting about a contact form. The form was bad. The form had always been bad. It collected name, email, and a message in a single 800px column, validated nothing in real time, surfaced the same generic success message whether you'd filled it in or not, and routed everything to the same shared inbox where it sat until somebody remembered.

I had a proper design. Inline validation. Smart field ordering. Context-aware microcopy. A success state that confirmed what would happen next and when. The pushback in the room came down to arithmetic. Two weeks of dev work, maybe three, for a contact form. The value isn't there for the time. We shipped what was there.

That arithmetic was usually right. Most "good UX" lived in Figma because the build cost rarely cleared the bar of what the work was worth. The contact form scene played out in some version, in some meeting, in basically every product team I worked on. Decent design ideas died on the way to engineering because decent design ideas cost real time, and real time was worth real money, and the maths didn't work.

That's the part that's gone. Building the contact form I had in my head used to take two weeks. It now takes an afternoon, because the tools are doing the part that used to consume the time. The same is true for a hundred other small decisions that used to die in cost conversations. Micro-interactions, error states, validation patterns, accessibility scaffolding, all the texture that separated "shipped" from "shipped well," generated, working, in front of you.

Here's where the argument gets honest, because the business view isn't wrong either. If you're a founder or a hiring lead, you've watched non-designers ship interfaces that would have needed a designer two years ago. Marketing pages, internal tools, dashboards, MVPs. The work isn't beautiful, but it works, and it costs nothing. That's a real shift. Anyone telling you the design layer is just as essential as it was is selling something. The floor moved up, and generic-good is now free.

AI gives you a foundation, and the wall is what that foundation doesn't include. It doesn't handle strategic thinking, personality, or justification well, and you can feel all three missing when you look at the output. It's why AI-generated work has a recognisable signature even when it's technically competent. Every contact form generated from a prompt looks like every other contact form generated from a prompt, because the model is averaging across what it's seen, and the average doesn't have a point of view.

With the right scaffolding, AI can get closer to all three. Carefully built skills, markdown files like CLAUDE.md, and custom instructions that hold a brand's voice and a product's strategic frame can pull the model away from the average and toward something specific. Setting up that scaffolding is design work in its own right. The skills, instruction files, and reference systems take taste and judgement to build, and the same to keep tuned, including noticing when the AI drifts and steering it back. The AI will drift.

And it hallucinates more than it one-shots. The "afternoon" version of the contact form is really minutes of generation followed by hours of finding the things that look right but aren't, untangling the inconsistent patterns, and making the small editorial calls that turn a passable form into one that fits the brand and the user it's actually for. That work is design. It always was. It just used to be wrapped inside the engineering bill, and now it's exposed.

Businesses are right to encourage AI in the design and build process, with emphasis on the assist. And they're right to raise their expectations of product designers because of it. It makes complete commercial sense, as long as it's used in moderation. For designers, that means embracing AI tools in the output while always steering the ship. Building out processes and finessing the ways of working with AI aren't nice-to-haves anymore. They're the work. Use it, and use it well, in moderation. A business isn't going to pay a designer to be a prompter. They should expect designers to be accountable for the output, whether AI was used to make it or not, whether it's code or anything else.