# NN/g Survey of 521 UX Practitioners: AI Saves Time, but 69% of Comments Are Negative > Source: > Published: 2026-10-24 00:00:00+00:00 AI makes UX work faster, and the people doing the work are only mildly happy about it. That is the headline of a survey the Nielsen Norman Group (NN/g) published on October 23, 2026. According to [NN/g's write-up](https://www.nngroup.com/articles/ai-in-ux-survey-26/?ref=explainx), 521 UX practitioners rated AI's effect on task time at 4.2 out of 5 (3 means no change), but rated how they feel about AI at work at just 3.57 out of 5. Among those who left an open comment, 69% expressed at least some negative sentiment. The gap between "faster" and "happier" is the interesting part, and it is relevant to anyone rolling out AI tools in a team. This post goes through the numbers, the splits by role and organization, what respondents said was dragging enthusiasm down, and why the sample needs careful reading. ## TL;DR: the survey in one table | Question | Answer (per NN/g) | |---|---| | Who was surveyed? | 521 UX practitioners (246 designers, 145 researchers, 21 product managers, 109 others), August 24 to September 11, 2026 | | Do they have AI access? | Over 99% have generative AI access; 59% use only an approved set of tools, 12% can use any | | Agentic AI use | 66% used at least one agentic tool; 85% at startups, 36% at nonprofits | | Top barriers | Output quality 54%, data privacy and security 50%, ethics or sustainability 34%; only 9% had none | | Time effect | 4.2 of 5 on average; most users said every listed task got faster | | Sentiment | 3.57 of 5 on average, mildly positive | | Open comments | 43% left one; 69% had some negative sentiment, 41% purely negative, 20% purely positive | | Role outlook | 65% expect their role to stay important or grow; 18% expect it to shrink or disappear | ## Who answered, and how much weight can the data bear? NN/g received 2,814 responses and removed incomplete ones, bot responses (all from LinkedIn) and people not currently working, leaving 521. Recruitment used a convenience sample from NN/g's practitioner-research panel, its newsletter, invite-only Slack communities and LinkedIn. The authors list their own caveats, and they matter: - A convenience sample may not represent UX practitioners overall. - Only 43% answered the optional open-ended question, and people with strong views are more likely to answer. - Product managers (21) and engineers (7) are small subgroups; the PMs were recruited through NN/g channels and likely know more UX than typical PMs. - Job security and role-change results are self-reported expectations, not observed outcomes. One methodological detail is worth noting for readers who follow AI evaluation: NN/g ran the sentiment analysis of open comments with two AI models, Claude Opus 5.5 and ChatGPT-6 Astra, using the same written rules. A researcher checked a random sample of 50 comments and reviewed every comment where the two models disagreed. An earlier, smaller survey gives a useful baseline. MeasuringU's analysis of the [2024 UXPA salary survey](https://measuringu.com/how-much-is-ai-used-in-ux/?ref=explainx) (444 respondents, mostly user researchers and designers) found 47% had tried AI and found some value, 20% tried it and were not impressed, and 37% said their company encouraged AI use. Two years on, NN/g's near-universal access and two-thirds agentic use show how fast the baseline moved, though the two samples and questions are not directly comparable. ## How much do UX teams use agentic AI? Access is almost universal, but freedom of choice is not: 59% use only an approved set of tools. The most popular tools were general assistants and chatbots, followed by AI features inside design tools. Agentic use is common but uneven. NN/g counted 66% of respondents as using at least one agentic tool, split between agentic AI products (48%) and AI coding agents or app builders (47%), with roughly 30% using both. A rough clay lump refined into a smooth vase, an image of iteration in agentic AI adoption for UX work ### Startups lead, government and nonprofits lag Agentic AI use was 85% among respondents at startups and 75% at small and mid-sized businesses, against 36% at nonprofits and 42% at government. The authors point to security reviews and cost in the slower sectors, and to light approval processes and small headcount at startups. By role, product managers and operations professionals used agentic AI most; content designers and researchers least, possibly because a general chatbot is often enough for syntheses and copy, and because agentic tools still need setup and troubleshooting. This lines up with the engineering-side picture in our post on [AI coding agents adding more code but not more finished software](https://explainx.ai/blog/ai-coding-agents-more-code-not-more-software-harvard-jellyfish-review-bottleneck-2026), where review capacity, not generation speed, was the limiting factor. ## Where does AI help, and where does it not? Respondents rated how often they use AI for each task and how it changed the time the task takes. Use and savings were highest on artifacts and data, and lowest when real users were involved. | Task | Share using AI often or always (among those who do the task) | Notes | |---|---|---| | Writing or editing code | 73% | Among the largest time savings | | Desk research | 69% | About 9 in 10 users said it was faster | | Analyzing own research data | 67% | About 9 in 10 users said it was faster | | Interviewing users | Never used by 61% | Smallest time savings | | Testing designs with users | Never used by 51% | Smallest time savings | Only 11% reported using AI interviewers, and the same share used synthetic users. The explanation is simple: AI can draft interview guides and test plans, but it does not shorten the sessions. Design evaluation is similar. AI can check a design against accessibility best practices, but critique depends on human goals and context that are hard to convey. ## Why is sentiment only 3.57 out of 5 if everything got faster? NN/g says this surprised them, and went to the comments. Of 226 comments, 155 contained negative sentiment, and 149 of those were coded with at least one of 12 reasons. The most common reason was unreliable output and the burden of checking and fixing it. ### The verification burden One respondent wrote that AI "can make 80000 mistakes in the time that it takes for someone to correct 5 of them." Others described fixing unusable prototypes made by colleagues and doing "double work" by reviewing others' AI output as well as their own. This is the same pattern we covered in [the AI verification tax](https://explainx.ai/blog/ai-verification-tax-rework-sage-workday-foxit-surveys-how-to-cut-it-2026): time saved at generation is partly spent again at checking. ### Mandates and inflated expectations Many respondents said their organizations mandated AI and tracked usage, which encouraged performative use. Lack of executive buy-in was the barrier chosen least often; several respondents described the opposite problem, constant management pressure. Others said leadership expectations of speed outran what output quality supports, and some received little training or guidance on which tools to use. A related finding from a manager-side survey is in our post on [Nielsen's 1,700-manager survey](https://explainx.ai/blog/monday-nielsen-survey-managers-ai-harder-to-evaluate-work-58-percent-more-agents-than-humans-1700-leaders-2026), which reported that AI-assisted work is harder to evaluate. Note that this is a different Nielsen from Nielsen Norman Group. ### Ethics and the environment 34% chose ethical or sustainability concerns as a barrier. Comments raised data-center environmental impact, the job market, AI-enabled surveillance and training on creators' work without consent. Some respondents described being required to use a technology they object to on moral grounds. ## What did the positive comments say? Nearly half of comments (49%) had some positive sentiment, most often about speed or productivity. Some said the saved time was reinvested into quality, and others said AI let them do things they could not before, such as coding or complex data work. The benefit looked greatest for solo practitioners and small teams: 5 of the 7 who mentioned working alone or on a very small team left purely positive comments (a tiny count, so treat it as a hint). Benefits and frustrations also came from the same people: of 61 comments mentioning being faster, 26 also expressed negative sentiment. NN/g's conclusion is that the benefits are real but uneven, clearest where practitioners control how they use AI and murkier under mandates, rising expectations or other people's AI output. ## Job security: who feels safe? 65% expect their role to stay as important or grow, and 18% expect it to shrink or be eliminated. Self-rated job security averaged 3.08 of 5 overall, with sharp splits: | Group | Job security (1 to 5) | |---|---| | Product management | 3.81 (n=21) | | Operations | 3.28 | | Content design | 3.04 | | Design | 2.99 | | Research | 2.94 | | Government or public sector | 3.86 | | Enterprise | 2.96 | | Small or mid-sized business | 2.87 | | Startup | 2.74 | Product managers rated security significantly higher than designers and researchers. The authors caution that the PM group is small and may be UX-adjacent, and note that interviewees said PMs were taking on more design and research work. By sector, government respondents felt most secure (3.86) and startup respondents least (2.74), and 31% of government respondents expected their role to stay largely unchanged versus 8% at enterprises. Note the irony: the sectors using agentic AI most feel least secure, and those using it least feel most secure. The survey cannot say whether one causes the other. Researchers were also more likely than designers to leave a purely negative comment (51% versus 33%), though once mixed comments were counted the gap (78% versus 66%) was not statistically significant. ## What should teams do with this? The survey is about UX, but the patterns travel. A few practical readings, which are our interpretation rather than NN/g's recommendations: 1. **Measure review time, not just generation time.** The 4.2 average speed rating describes the task as respondents perceived it; the comments suggest checking and fixing is where the time goes back. 2. **Mandates with usage tracking invite performative use.** If you track AI usage, expect output that someone else must clean up. 3. **Match the tool to the task.** AI saved the least time in sessions with users. Do not expect uniform gains across a workflow. 4. **Give guidance, not just access.** Respondents with approved tools but no training described carrying the learning burden themselves. 5. **Treat security and privacy as first-order.** Half of respondents named it a barrier, and the sectors with the strictest review adopted agentic AI the least. For a broader view of how builders are handling the checking problem, see [shipping AI-written code you do not understand](https://explainx.ai/blog/shipping-ai-written-code-you-dont-understand-ask-hn-meta-engineer-playbook-2026), and for another self-reported workplace dataset, [Gallup's survey of US teachers](https://explainx.ai/blog/gallup-teachers-ai-59-percent-used-64-percent-more-time-mattering-survey-2026). ## Bottom line NN/g's survey adds a clean data point to a pattern showing up across professions: AI does reduce task time, and people still feel ambivalent because the cost shows up elsewhere, in verification, mandates, expectations and job worries. The sample is self-selected and small in places, so treat the exact percentages as indicative of this panel, not of UX as a whole. The consistent message across the comments is that how AI is introduced matters as much as which tool is introduced. ## Related reading *Figures are accurate as of October 24, 2026 and come from NN/g's published write-up; check the source for corrections.*