Last updated: 26 July 2026.
AI Can Read a Thousand Users. It Can’t Understand One.
AI can transcribe a hundred interviews and summarize a thousand survey responses before your coffee’s cold. Most research teams have noticed: 58% now use AI tools, a 32% jump in a single year (Maze, 2025). But speed isn’t the same as understanding. AI reads the data. It still can’t sit with one confused user and feel why they’re stuck.
This guide is a workflow, not a tool dump. It walks through how to use AI across the UX process, research, synthesis, testing, and iteration, and where the human part has to stay human. For the ranked tool list, we point you to our guide to the best AI tools for user research so this piece can focus on the how. The short version: use AI to scale the analysis, not to skip the understanding.
Key Takeaways
- 58% of research teams now use AI, up 32% in a year (Maze, 2025), mostly for analysis and transcription.
- AI is fastest at the volume work: 74% of teams use it to analyze research data (Maze, 2025).
- It can’t replace real users. NN/g found AI “synthetic users” too sycophantic to stand in for real research (NN/g, 2024).
- Use AI to process data faster, then spend the time you save on talking to and understanding real people.
Table of Contents
What Does AI Actually Change in UX Work?
It changes where your hours go, not what UX is for. In Maze’s 2025 survey of 800 practitioners, 58% used AI tools, up 32% from the year before, and most of that use lands on analysis and transcription (Maze, 2025). AI eats the slow, mechanical middle of research. The thinking on either side stays yours.
The clearest way to see the split is to line up what AI genuinely does against what it can’t. One column is why teams adopted it. The other is why the job still needs you.
| AI can | AI still can’t |
|---|---|
| Transcribe and tag hundreds of interviews in minutes | Sit with one confused user and feel the frustration |
| Summarize survey data and surface recurring themes | Decide which theme actually matters for the product |
| Draft research plans, screeners, and discussion guides | Ask the unscripted follow-up that unlocks the real insight |
| Spot patterns across a large dataset | Understand why one person behaves against the pattern |
One caution before you lean in. AI summaries read as confident and complete, which makes them easy to over-trust. A model will happily invent a tidy theme from messy data, or flatten a strong minority view into a footnote. Treat every AI summary as a first draft of the truth, then check it against the raw sessions, not as the finding itself.
How Do You Use AI Across the UX Process?
Map AI to the stage, not the whole job. The process still runs research, synthesize, test, iterate, and AI helps most in the middle two. Teams already lean on it heaviest for analysis (74%) and transcription (58%), with 57% reporting faster turnaround (Maze, 2025). The video below is a working designer showing exactly where AI slots into that flow.
Stage 1: Research (AI drafts, you talk to people)
Use AI to prepare, not to skip the conversation. It can draft your screener, generate a discussion guide, and summarize prior studies so you walk in prepared. But the interviews themselves need a human. The unscripted “wait, why did you do that?” is where real insight lives, and no model asks it for you. There’s a quieter risk too. If AI writes your discussion guide, it drifts toward the obvious questions everyone asks. Edit it. Add the weird, specific question only someone who knows the product would think of, because that’s usually the one that pays off.
Stage 2: Synthesize (AI’s strongest stage)
This is where AI earns its place. Transcription, tagging, and first-pass theme clustering across dozens of sessions used to eat days. Tools like Dovetail, Notably, and Looppanel now do the grunt work in minutes. Your job shifts to the harder question: of the themes it surfaced, which ones actually matter, and which are noise the model over-weighted? One convention that helps: let AI tag first, then rename its tags in your own words. Its labels come out generic (“navigation issues”); yours carry the specific nuance (“users miss the back button on step 3”).
Stage 3: Test (AI assists, real users decide)
AI can help you write test tasks, recruit, and summarize sessions. What it can’t do is watch. NN/g is blunt about this: AI tools “are not actually watching” where people look or hover, so they shouldn’t moderate usability tests that depend on behavioral observation (NN/g, 2024). Run the test with real people. Let AI clean up the notes afterward. A safe division of labor: let the model draft the task scenarios and the post-session summary, and keep a human in the room to read the hesitation, the backtracking, and the body language a transcript never captures.
Stage 4: Iterate (AI speeds the loop)
Once you know what to fix, AI shortens the loop. It drafts revised copy, generates layout variations to react to, and helps prioritize the backlog against your findings. The trap is generating so many options that you stop deciding. Ask for three variations, not thirty, then pick one and move. When you move from research into the visual layer, our how-to on AI for UI design picks up where this leaves off, at the interface itself.
A worked example: 40 interviews, one afternoon
Say you’ve just run 40 user interviews about a checkout flow. The old way: a week of re-watching recordings and sticky notes. The AI way: drop the transcripts into Dovetail or Notably, and in an afternoon you have them tagged, clustered, and summarized into candidate themes.
Then the real work starts.
The model hands you eight neat themes. But it flagged “users want more payment options” as the top issue, when what you actually heard, in the pauses and the sighs, was that people didn’t trust the page enough to enter a card at all. AI counted the words. You caught the hesitation.
So you cut two themes it invented, merge two it split, and rewrite the top finding around trust. AI saved you the week of transcription. The judgment call that reframes the whole project is still yours, and that’s the part that ships a better product.
Which AI Tools Fit the UX Process?
The ones built for research, not the ones built for pixels. A quick note on naming: Galileo AI is now Google Stitch (acquired by Google in 2025) and Uizard was acquired by Miro, and both are UI/visual tools, so we cover them in our AI for UI design guide, not here. For the UX process itself, these are the ones that matter. For the full ranked list, see the best AI tools for user research.
| Tool | Best for | Stage |
|---|---|---|
| Maze | Unmoderated testing with AI analysis and survey generation | Test |
| Dovetail | Research repository with AI tagging and insight summaries | Synthesize |
| UserTesting | Moderated and unmoderated testing with AI session summaries | Test |
| Notably | AI-forward theme extraction and synthesis | Synthesize |
| Looppanel | AI note-taking, transcription, and thematic tagging | Synthesize |
| ChatGPT | Drafting plans, screeners, and first-pass analysis | Research / Synthesize |
Pick by stage, and don’t buy five tools when two cover your workflow. Most teams get the biggest win from one synthesis tool and one testing platform. For the wider creative stack, our overview of AI for design ties the UX and visual sides together.
Can AI Replace UX Designers or Real Users?
No, and the sharpest evidence is about fake users. When Nielsen Norman Group tested AI “synthetic users” against real prior studies, it found them sycophantic and shallow: they “seem to care about everything,” and its guidance is explicit, don’t use synthetic-user research as a replacement for real-user research (NN/g, 2024). A model trained on averages can’t reproduce the one person who breaks the pattern.
Designers feel the limit too, and the data holds a real tension. 81% of designers say AI dulls creativity, the highest of any creative field (Exeter, 2025), yet 90% say design is at least as important as it was before AI (Figma, 2026). Read together, that’s not fear. It’s practitioners saying the craft matters more precisely because the easy parts got automated.
The most authoritative voice on AI participants sums it up in a few minutes. Nielsen Norman Group’s short explainer below is worth watching before you let any tool tell you what “users” think.
Frequently Asked Questions About AI for UX Design
What is the best AI tool for UX research?
It depends on the stage. Maze and UserTesting lead for testing, Dovetail and Notably for analysis and synthesis, and ChatGPT for drafting plans and first-pass themes. With 58% of research teams now using AI (Maze, 2025), the winning move is matching the tool to the research stage, not picking one winner.
Can AI replace UX designers or researchers?
No. AI handles the volume work, but NN/g found AI synthetic users too sycophantic to replace real research (NN/g, 2024), and 90% of designers say design is at least as important as before AI (Figma, 2026). AI shifts where your time goes; it doesn’t remove the judgment.
Can AI run user research on its own?
Not reliably. AI can transcribe, tag, and summarize, but it can’t observe real behavior or feel a user’s confusion. NN/g’s guidance is explicit: don’t use synthetic users as a replacement for real-user research (NN/g, 2024). Use AI to scale analysis, then talk to real people.
How much time does AI actually save in UX research?
Mostly in analysis and transcription. In Maze’s 2025 survey, 74% of teams use AI to analyze research data and 57% report faster turnaround (Maze, 2025). The savings are real, but they land in processing the data, not in understanding the people behind it.
Which UX tools are AI-powered?
Maze, UserTesting, Dovetail, Notably, and Looppanel all have AI features for testing, transcription, and analysis. Worth noting: Galileo AI is now Google Stitch and Uizard was acquired by Miro. Both are UI/visual tools, covered in our UI guide, not UX-research tools, so don’t confuse them for research platforms.
The Bottom Line
AI changed UX by taking over the slow middle: transcribing, tagging, summarizing. That’s real, and 58% of research teams already use it (Maze, 2025). But UX was never about processing data. It’s about understanding people, and a model trained on averages can’t do that for you.
So let AI carry the volume, and spend the time it gives back on the part that matters: talking to real users, catching the hesitation in the room, and deciding what’s actually worth building. For the tools, start with the best AI tools for user research. For the interface side of the work, our how-to on AI for UI design is the natural next read.
Richard Johnson
Richard Johnson is an AI specialist with over five years of experience guiding large organizations through AI adoption, across more than 100 customers. He founded CognitiveFuture to research and compare AI tools across design, development, writing, research, voice and business, cutting a crowded, fast-moving market down to the right choice for the job in front of you.
