AI design tools can cut task time by about 20% — but does that hold for every kind of design work?

This explores whether the roughly 20% time savings from prompting an AI design tool holds across different kinds of design work, or only for certain kinds. The collection has one direct measurement, and it covers only part of the question. In a randomized trial of 100 designers and product managers, Figma Make access cut completion time by about 20% on *structured* tasks, and product managers gained more than designers did Does Figma Make speed up design task completion?. So the speedup already varies by who is doing the work. Nobody in the trial ran open-ended, exploratory design, so the result can't yet be extended to it.

Adjacent evidence suggests that task type matters. A benchmark of 23 prompt styles across 12 models found that which prompts help depends on task structure and model tier, not on any universal best practice. Step-by-step reasoning even lowered accuracy in the strongest models Do prompt techniques work the same across all LLM tiers?. Another study found that for simple, well-defined questions, going straight from question to answer beats elaborate reasoning. The best prompt depends on the kind of question, not just the broad task category Why do some questions perform better without step-by-step reasoning?. A design brief with a clear spec looks like that easy, well-defined case, and the study's structured tasks probably were.

There is also a hint that models treat structured and creative work differently at a deeper level. In multi-task training, structured domains push a model's output variety down while creative domains push it up, and the order in which you train them changes how well each survives Does training order reshape how models handle different task types?. That research is about training, not design tools. Still, if the two kinds of work pull in opposite directions inside the model, a speedup measured on one kind is a shaky basis for predicting the other.

The likeliest reason open-ended design would gain less is iteration. Prompt refinement works by steering the output toward what you already expect, so the result is a co-production of you and the model How much does the user shape what a model generates?. When the target is clear, a few rounds converge fast. When you don't yet know what you want, each round can also shift the goal, and that eats the savings. This is an inference from the corpus, not a measured result.

The trial's outcome measure is also narrow. Completion time says nothing about whether the design was better. A production case shows how a metric can improve while the underlying quality doesn't: a prompt lifted a pass rate from 23.1% to 80.0% by learning what the judge liked, and precision didn't move Can prompt optimization accidentally teach judges to reward the wrong signals?. The Figma trial's gains in ease and perceived usability are real, but they are also perceptions. The collection leaves open whether the speedup holds for open-ended work, and whether faster means better in either kind.