Seeing the amount of AI tools that have shipped this quarter, all promising the same trick: point at a brand and generate the interface and skip the design system entirely, none of it worries me, and that’s not bravado.
Every one of those tools, the genuinely impressive ones and the gimmicky ones, still has to answer the exact same question the moment it tries to render a single pixel of somebody’s brand: what does this actually look like, expressed as something a machine can read. That question was never going away, and it was never going to belong to whichever tool got there first. It belongs to whoever already had the answer sitting there, structured and waiting, before the tool showed up asking.
That’s the part I keep having to explain, because it runs against the instinct everyone in this industry has right now, which is to treat every new release as a threat clock. Something ships, and the panic is immediate: does this replace us, does this replace the system, do we need to rebuild everything to keep up. I understand the instinct. I just don’t share it, and it’s not about being unbothered by nature. It’s that the layer I work at was never built to compete with what’s coming. It was built to feed it.
The question every new tool still has to ask
Here’s what I mean by that, concretely. Whatever generates an interface, whether it’s a person in Figma, a script pulling from a component library, or a model producing a screen from a sentence, has to pull from something more specific than “make it look nice.” A color. A type scale. A rule for exactly how much padding sits around a button before it reads as wrong to someone who’s seen the brand a thousand times. Even a model dreaming up a layout from nothing is sampling from a distribution of what buttons and spacing and color look like in general, and the more precisely that distribution gets narrowed to one brand’s actual rules, the less the output drifts toward generic. Vague input gets you a plausible guess. Structured input gets you the brand.
That’s the subatomic layer, and it’s the same layer whether the thing consuming it is a junior designer, a build script, or whatever ships next quarter under a name none of us have heard yet. A2UI, agentic design tools, some fourth thing nobody’s coined a term for. It genuinely does not matter which one wins. All of them need the same unglamorous answer: not a vibe, not a PDF of brand guidelines, but values, named and related to each other, sitting one level below the components everyone’s already used to thinking about. Get that layer right once, and you’ve built something that doesn’t care what ends up standing on top of it. Get it wrong, or skip it because it feels too invisible to prioritize when there’s a client in the room, and every new tool just becomes one more thing you have to manually keep from drifting off-brand, forever, by hand.
What it looks like when it actually works
I’ve watched smaller versions of this play out for a couple of years now. One system I’ve had live held a brand’s base typography steady for two years straight while everything built on top of it kept moving. Then a campaign needed something louder, so we layered in a second typeface just for that push, sitting on top of the base without touching it, and stripped it back out once the campaign wrapped. The foundation never got redesigned to make room for the loud font, and it never got redesigned to remove it either. It just absorbed both and kept running underneath, exactly as invisible the whole time as it’s supposed to be.
The same pattern shows up at a different scale when a brand adds a market, or a channel, or shifts a color system for an accessibility requirement nobody saw coming a year out. None of those are redesigns. They’re additions at the edge, sitting on top of rules that were never load-bearing for that particular decision in the first place, which is the entire point of drawing the line where you draw it. You’re not trying to predict every future need when you build this layer. You’re trying to make sure the layer doesn’t care what the future need turns out to be. Those are two completely different problems, and confusing them is how teams end up rebuilding their foundation every eighteen months instead of building on the same one for years.
That’s the whole trick, and it’s a much less exciting trick than it sounds like when you say “AI-native design system” out loud in a pitch room. Nobody claps for “we didn’t have to change anything”; that’s the actual deliverable, quietly, every time it happens again.
Why I’m not racing the roadmap
I should be honest that this isn’t a comfortable position to hold right now, because half of what gets funded in this industry is the promise of staying ahead of the next tool. Every deck wants to be the one that shows the new capability before a competitor finds it. Every pitch wants a line about being on the frontier. There’s real pressure to spend your energy watching the roadmap of whichever AI lab is loudest that quarter, and to angle your whole strategy around guessing what they ship next, and to redo the strategy again three months later when they ship something else.
I’ve stopped playing that game, and not because I think the tools don’t matter. They obviously matter, and some of them are genuinely great. I’ve stopped playing it because it’s a race you cannot win by definition. You don’t control their timeline, you don’t get a preview, and they are not returning your calls. Every hour spent trying to predict the next capability is an hour not spent making sure your own brand is legible enough that whatever does show up can read it correctly the first time, without a quarter of cleanup trailing behind it. One of those bets compounds, quietly, in the background, whether you’re paying attention to it that week or not. The other one resets every time somebody ships something new, which lately is about every six weeks.
So the actual strategic question was never “what’s coming.” It was always “are we ready for whatever it turns out to be,” and those two questions point you toward completely different work, with completely different weeks.
They all have to stand on something
Underdrawing isn’t placing a bet on which tool wins next. It’s placing a bet on the floor every one of them eventually has to touch. Get the tokens right, get the call on what’s locked and what’s left open right, and the next tool that shows up doesn’t threaten anything. It just becomes one more thing that has to stand on something solid, the same way the last one did and the one before that.
They all have to. Every single one, no matter how much smarter the next version is than the last, no matter how confident the launch clip sounds. That was never a limitation on what’s coming. It’s the only part of any of this that was ever going to hold still long enough to build on.