The 70/30 Rule

Where this came from

I use AI every day. I have for two years. This isn't an article about resisting it. It's about the line I keep bumping into when I let it do too much.

The line shows up everywhere. In workshops I run, in client work, in side projects, in code I'm shipping. Every AI output hits a ceiling. The research summary misses the subtle insight. The layout is technically correct and emotionally flat. The component compiles and is brittle in ways you can't articulate from the file alone. The copy communicates without connecting. There's a moment, looking at any of it, when you realise the work isn't bad but isn't going to land either, and you can't quite say what's missing because what's missing is the thing that makes you bother in the first place. The ceiling lands around 70%, consistently enough that I started calling it a rule.

There's a comfortable take that AI is just a tool, like Figma was a tool, like Sketch was a tool. That's true for the 70%. It's the most misleading thing you can say about the 30%.

Where the split falls

70% is the average. It's a heuristic, not a measurement. Pareto, not physics. The interesting question is which tasks sit far above it, and which fall far below.

The shape is always the same. AI is fast and wide: it generates, summarizes, explores. Humans are slow and narrow: they choose, refine, decide. The 70/30 isn't a number to memorize. It's the line between fast-and-wide and slow-and-narrow, and where that line falls depends on what you're making.

The temptation to stop at 70

Using AI for the first 70% isn't the risk. That's just efficiency. The risk is shipping AI's version of the last 30%, because AI's version is almost always almost there. The layout works. The copy makes sense. The flow is logical. Nothing is wrong with it. That's the problem.

Almost-there is where craft goes to die. The spacing that's technically correct but visually wrong. The transition that's smooth but emotionally flat. The button label that's grammatical but lifeless. The component API that works for today and fights you the moment requirements shift. None of it is broken. None of it is right either. And it's the only part users, and the next developer, actually feel.

Taste is a debt

When I say the last 30% is where craft lives, I'm not really talking about polish. I'm talking about taste. They get confused, but they're different things.

Polish is surface work, the kind you can name and describe: the spacing, the radius, the easing curve. Taste is messier. It's the discipline of knowing what to keep and what to throw away, and most of it can't be put into words. You can only show it by doing.

Here's the thing nobody is saying out loud: taste isn't a skill you have. It's a debt you've paid. You develop taste by making bad work, recognizing it as bad, and trying again. That's the whole process. There is no other one.

AI is good at the part where you make work. AI can also judge work. There's a whole field of using AI to evaluate AI. But AI judges by the patterns of what's already been called good. That's the math of how it was trained. A deliberate departure from those patterns scores worse on the only distribution AI knows. Taste is the eye that recognizes when a departure is the work, and when it's a mistake. It's the only judgment that matters in the 30%.

Which means: a designer who lets AI do 70% of the work, without having paid the taste debt themselves, will ship the AI version every time. They won't see the difference. They'll think they're doing the job. The output will be there. Whether it's good or not is a question that requires taste to even ask.

The opposite failure mode matters too. The designer who refuses to use AI, who still does the first 70% by hand because that's where they feel professional, gets outpaced by the one who uses AI aggressively and knows where to stop. Both extremes lose. The 10/90 designer falls behind. The 90/10 designer ships fast and unrefined. The middle is harder than either extreme, which is why most people pick a side.

How it changed the way I work

Recent client project. AI gave me twelve dashboard layouts in under an hour. Two years ago, that would have been three days. The interesting part wasn't picking the best one. It was deleting eleven of them, and then deleting half the elements inside the one I kept.

AI-generated design has a very specific weakness: it can't bring itself to leave things out. Everything that might be useful is there, everything that might matter has weight. The hours I'd saved went into removing things: the second CTA, the redundant label, the chart that explained what the heading already said. The kind of work that, when it's done right, looks like nothing happened.

That's the move. Every hour AI saves on the 70% gets reinvested in the 30%. Most of it goes into saying no to what AI offered: to layouts, to components, to clever code I'd regret in six months. Not shipping more, faster. Shipping less, better.

When the number moves

It will. The 70 will become 80. Maybe 90. The percentage is the least interesting part of the rule. What stays is the shape: AI generates, humans judge. The line between divergent and convergent doesn't move with model versions. It moves with what you're making.

In two years, AI's first try will land where my twelfth lands today. The hours I spent deleting things? Still mine.

Maybe I'm wrong. Maybe AI will develop the kind of judgment that decides what's bad about its own output. If it does, the rule disappears. And so does most of the job. That's not the bet I'm making.

The paradox

Everyone is working with the same AI now. You'd think that would flatten the field, and at the 70% mark it does: outputs converge on competent, balanced, fine. What happens at the 30% mark is the opposite. The gap between the people who can see past fine and the people who can't is suddenly very visible. Craft isn't threatened by AI; it's exposed by it. And the designers who haven't paid the taste debt are about to look very obvious.