Figma's own trial found prompt-to-design saved PMs the most time

Most claims about AI prototyping tools are demos. A paper posted on arXiv on 22 September 2026 is one of the few controlled trials: product designers and product managers, randomly assigned to work with or without a prompt-to-design tool, timed on the same tasks. The result is a real saving, and the conditions under which it was measured are the most useful part for anyone who prototypes from a spec.

The trial was run by Figma, on Figma's own product. Read it as vendor research — careful vendor research, with an external firm doing recruitment and moderation, but vendor research.

The study

Four researchers at Figma designed a between-subjects randomised trial of Figma Make, the company's tool for generating designs and prototypes from conversational prompts. The research firm MeasuringU recruited 50 product designers and 50 product managers, each with at least two years in role, who used Figma at least monthly and had never worked for Figma. Within each role, half were randomly assigned to use Figma Make and half to edit manually.

Everyone did the same three edits to a social-media app design, in the same order:

1. Visual styling — convert the design to dark mode.
2. Interface modification — add a "Help and Support" option to a settings menu.
3. Interaction prototyping — build a comment flyout that opens when the comment icon is clicked.

For each task, participants received the starting design and a static image of the finished target. A moderator stopped the clock once the submission matched the reference. Treatment participants got a five-minute primer and could use Figma Make plus manual edits; they sent on average 2.9, 2.6 and 5.2 prompts on the three tasks.

What they found

About a fifth faster overall. Among participants who completed the tasks, access to Figma Make cut completion time by 20% when the three tasks were weighted equally, and by 27% — roughly 15 minutes — on cumulative time. Task 2 was 27% faster and Task 3 32% faster. The dark-mode task showed no significant difference.

Product managers gained most. PMs with the tool saved 22 minutes 46 seconds across the three tasks, a 35% reduction. Their task-ease ratings rose 37% and their usability scores 32%. Without the tool, PMs took considerably longer than designers; with it, the authors say, that disadvantage almost disappeared.

Designers gained on the hardest task only. Designers' only clear saving was on the interaction prototype: 7 minutes 41 seconds, or 26%. Their overall saving of 17% was marginally significant, and their ease and usability ratings did not change significantly.

More people finished the interaction prototype. Overall, 73% completed Task 3. Access to Figma Make raised the probability of completing it by 19.9 percentage points, driven by PMs.

People rated their own work higher — at the extremes. Participants in both groups mostly put their study designs at about the same or somewhat worse than what they make outside the experiment; the conditions parted company in the tails. Treatment participants were more likely to rate their designs as much better than what they could make outside the study (20% versus 2%). That difference was significant for PMs and not for designers. 78% of treatment participants somewhat or strongly agreed the tool would increase their ability to contribute to design. Asked whether they trusted it to interpret their intentions correctly, 42% somewhat agreed and 32% were neutral.

The caveats, from the paper itself

The authors list five limitations, and they are specific.

• Only completed tasks were timed. Because the tool also changed who completed tasks — especially PMs on Task 3 — the groups being timed may differ. The authors say the time results should be read alongside the completion results.
• The tasks had answers. In the authors' words, the structured tasks did not capture "the ambiguity, need to encompass stakeholder feedback, and changing requirements that are common in professional design projects."
• Moderators judged the match. Human variability in checking submissions against reference images may affect results.
• The tool sometimes failed. Stochastic rendering failures added time for treatment participants, which the authors say may mean the savings are underestimated.
• A specific population. 100 regular Figma users, with PMs as the only non-design role. The results may not hold for newer users, other roles or other prompt-to-design tools.

Two more points are worth adding. Design quality was self-assessed; the authors recommend expert ratings with predefined rubrics in future work. And they flag the question nobody has answered yet: "whether time saved during initial production is offset by stakeholder review or later iterations."

Why this matters if you prototype from specs

The headline will travel as "AI makes design 20% faster". The more useful reading is about input.

Every task in this trial came with the answer attached. A reference image of the finished design settles what the result has to look like: the palette, the new menu item, the flyout open on screen. Little is left to decide — and notably, the authors supplied that target as a static image rather than an editable frame, so its internals could not simply be copied. The task a still image could least fully pin down, the interaction, was also the longest and the least often finished. Under those conditions, a prompt-to-design tool saved real time, most of all for the people least fluent in the design tool. That is a meaningful result. It is also close to the easiest case.

Real prototyping starts from a spec that decides some things and is silent on others. The prototype then has to guess, and as we have argued before, a plausible guess is harder to catch than a blank space. This trial did not measure what happens when the target is a paragraph rather than a picture, and its authors say so. The gap between "the spec decided" and "the prototype guessed" is the subject of the prototype examples post that runs alongside this briefing.

The self-rating result is a review problem. PMs with the tool rated their own designs much better than what they could have made outside the study, and nobody independent checked. When a PM can produce a convincing interaction prototype in minutes, the scarce step is no longer making it. It is someone confirming that it shows what the spec meant — and noticing what it invented. That is why the trial's open question about stakeholder review is the right one.

Time saved is half the ledger. Our earlier briefing on how teams measure AI by time saved made the case that speed is the half of the result that looks good. This study is more careful than most: it reports completion, ease and self-rated quality alongside time. It still cannot see what happens after the prototype leaves the room.

What to take from it

Prompt-to-design tools can save PMs meaningful time on well-defined UI changes. That is a fair reading of a careful, if vendor-run, trial.

The narrower lesson for spec writers: the savings were measured on tasks where the target was fully specified. The closer your spec gets to that — named screens, explicit states, concrete content — the closer your prototyping gets to the conditions this trial measured. Where it falls short, the tool will not slow down to ask. It will fill the gap.

FAQ

What was tested? Figma Make versus manual editing, for 50 designers and 50 PMs, on three timed UI tasks with reference images.

The result? About 20% faster on average among completers, 35% for PMs on cumulative time.

Who ran it? Figma researchers, with MeasuringU handling recruitment and moderation.

Biggest caveat? Every task had a fixed visual target, so ambiguity, stakeholder feedback and changing requirements were absent by design.