“I now sometimes find myself trying to quickly fix completely unusable prototypes made by a product manager.” That is one respondent in a Nielsen Norman Group survey of 521 UX practitioners, published on 23 October 2026. Another wrote that the people who care about what a team makes “end up doing double work, having to review and fix the slop of others”. NN/g gives no job titles for either, so call them respondents. The tension is plain enough: AI makes each person’s task faster, but the saved time often becomes work for somebody else, and the usual way of measuring AI only times the first person.
What the trials time, and what they skip
Start with the best evidence that AI saves time in design. In a randomized controlled experiment on arXiv , 50 product designers and 50 product managers each completed three standardized design tasks, with or without Figma Make. Access was associated with “approximately 20% shorter completion times”. The gains were “larger among product managers”, and for designers they “may be task dependent”. We covered that role split in an earlier piece on the trial .
Look at what the abstract measures: one person, one task, a stopwatch. It reports no quality measure and no downstream rework measure. That is a fair design for the question it asks. It is a blind one for the question leadership will ask next.
NN/g’s survey ran from 24 August to 11 September 2026. Respondents rated AI’s effect on task time at 4.2 out of 5, where 3 means no change. So the speed is real in their experience too. Yet their overall feeling about AI came in at 3.57. The gap is in the comments, where NN/g writes that “the time saved rarely stayed with them”: it was absorbed by expectations to produce more, or spent cleaning up other people’s AI-generated work. Its own summary line is the sharpest one: “A task made faster for one person can create work for another.”
Where a saved hour goes
Here is a mechanism you can run in your head. A PM finishes a prototype task faster with Figma Make, and a trial books that as a saving. The prototype then reaches a designer, who has to rebuild it before anyone can test it. None of that rebuilding appears in the trial. The saving is counted once. The cost is not counted at all.
That is the stake for a product designer whose PM can now produce a clickable prototype in an afternoon. The designer still has to make it usable. Their time doesn’t show up in the time-saved figure that leadership uses to set next quarter’s expectations.
The expectation half of the story has an independent echo. A UC Berkeley Haas ethnography at a roughly 200-person US tech company found employees took on a broader scope of tasks, including work that “previously would have belonged to someone else”. Researcher Xingqi Maggie Ye said “expectations can gradually reset, and what was once extra effort becomes standard performance.” That study isn’t about design, and it is still in progress. It shows the mechanism, not the size of it.
A metric that times individual tasks cannot see work that moves from the fast person to the person downstream.
The ledger nobody has opened
Now the leap, which is ours and not the sources’. Figma’s trial measured individual completion time. NN/g coded free-text comments about time lost to expectations and to cleanup. No source measures net team time, or how much of a PM’s saving reappears as a designer’s rework. Put the two side by side and the reasonable conclusion is that per-person task time probably overstates team gains from AI in design. The instruments in use cannot register work that moves from the fast person to the person downstream.
The same goes for usage tracking. NN/g ties tracked usage to performative use, and one respondent describes “the constant push from management to use it more and more”. A dashboard that counts logins is scoring compliance, not time.
What the counterevidence allows
NN/g itself warns that its comment-based findings “should be interpreted with some caution”. Respondents with strong views were more likely to comment. The sample was drawn through NN/g’s own channels. Two AI models coded the sentiment. Of 226 commenters, 69% voiced some negative sentiment, which a self-selected pool may inflate.
The survey also has clearly happy voices: “With AI I can complete work on most days within hours now. This frees up so much more valuable time.” NN/g adds that benefits are clearest for practitioners who control how they use AI, and only 12% could use any tool they chose. The Berkeley study complicates the management-pressure reading too, since it found scope growing mostly voluntarily, without any mandate.
None of this rescues the per-task number. It sharpens the claim. Some people keep their saved hours, some lose them, and a stopwatch on one desk can’t tell which. Until someone times the handoff as well as the task, the saving is the half of the ledger that looks good.
