The last 20% is the whole point

Designer Fund's 2026 AI in Design report found that half of all designers have shipped AI-generated code to production. Weekly AI usage jumped from 54% to 91% in a year. And the top complaint has not changed: unreliable output quality.

That number is worth sitting with, because it contradicts the hype. Generation is solved. Judgment is not. AI produces a competent first pass in seconds, and then the product stalls in the last 20% — the part where hierarchy, edge cases, and copy actually get decided. This is the same shift I described in The Founder Floor: your job moved from writing everything to reviewing what the machine produces and applying taste at the gate.

Paul Bakaus, who built the Impeccable design-skills system, argues against one-shot AI design in an interview with Latent Space. His model is explicit: let AI build the first 80% fast, then a person owns the final 20%. He refuses to add an auto mode. "There is no auto," he said, "and there will be no auto."

Why is the last 20% the hard part?

The last 20% is the only part where taste has a job to do. The first 80% is reconstruction — a model averages every interface it has seen and assembles a competent default. The last 20% is decision: which hierarchy matters, where delight goes, what gets cut.

Researchers at the UC Berkeley School of Information observed 13 design sessions and documented six ways prompt-mediated design breaks down: a vocabulary gap, an execution gap, a convergence trap toward generic polished styles, a tacit ceiling, an authorship disconnect, and risk-reward freeze. Every one of those breakdowns happens in the last 20%. None of them are about the model being too weak.

What does the 80/20 split look like in practice?

It looks like a founder who can spin up a working prototype in under 20 minutes and still spends three days making it feel intentional. The Designer Fund report shows why: the average designer now uses 7 AI tools regularly, up from 3 a year ago, and 65% are taking on more product and engineering responsibility. Everyone can generate. Almost nobody can finish.

The first 80% gets you structure, spacing, and a layout that does not embarrass you. The last 20% is the accessibility pass, the responsive edge cases, the empty states, and the copy that reads like a person wrote it. That is not a bug in the tools. It is a division of labor — and the human side of the split is the side that requires taste.

How do you own the final 20%?

You make taste into process instead of a feeling. Start with an eval set — a list of real inputs with the outputs you consider good — which is exactly what I wrote about in Your AI eval set is your taste, made measurable. Then apply three habits:

1. Critique every output against your standard before accepting it. The default is a polished average, and accepting it is how generic products ship.
2. Write the guardrails down. Spacing, hierarchy, what the product should not feel like. Agents follow explicit constraints; they will never guess your point of view.
3. Cut harder than you generate. The last 20% is mostly deciding what does not earn a place.

Will AI close the gap?

No — and the evidence points the other way. Only 5% of leaders in the Designer Fund survey said they are placing less emphasis on execution quality. As generation got cheaper, taste and judgment became the hiring priority. Output quality is the single biggest barrier to AI adoption, which means the interface layer is where products win and lose.

AI makes the last 20% faster to iterate. It does not make it cheaper to decide. That is a structural gap, not a temporary one. The first 80% is now free, which means everyone has it. The last 20% is where your product stops being a category and starts being yours. That part is still — and should stay — your job.