Using AI for UX Research: The Workflow I Actually Use

Here’s what actually works when you use AI for UX research — and what’s just expensive procrastination.

I resisted using AI for UX research for longer than I probably should have.

Not because I’m anti-AI or think it’s going to steal jobs. But because most “AI research tools” seemed designed to help you skip the actual work of understanding users. And I’ve seen what happens when teams use research as an excuse instead of a tool.

But over the past year, I figured out which parts of UX research AI actually helps with, and which parts are just marketing promises that waste your time.

This isn’t a comprehensive review of every AI tool for UX research. It’s what I actually use in my UX design work, what I’ve stopped using, and why.

The short version: AI is excellent at the boring parts of research. It’s terrible at the parts that actually matter.

What AI Is Actually Good At (The Boring Stuff)

Let’s start with where AI for UX research tools genuinely save time:

1. Transcription (Finally, Something That Just Works)

What I use: Otter.ai or Descript

Why it works: Transcribing a 45-minute user interview used to take 2-3 hours. Now it takes 5 minutes and costs nothing.

The catch: You still need to clean it up. AI misses context, mishears technical terms, and doesn’t capture tone or emphasis. But getting 90% of the words down automatically beats typing everything manually.

My workflow:

• Record the interview (with permission, obviously)
• Upload to Otter immediately after
• Review the transcript while listening to key moments
• Mark sections worth pulling quotes from
• Export to Google Doc for synthesis

Time saved: About 2 hours per interview. That’s real.

2. Initial Theme Identification (But Don’t Trust It)

What I use: Claude or ChatGPT with custom prompts

Why it works: AI can scan through 10 interview transcripts and spot recurring patterns faster than I can. It’s pattern matching, which is what AI does well.

The catch: It finds patterns that aren’t meaningful and misses context that matters. You absolutely cannot skip the manual review step.

My workflow:

• Feed transcripts into Claude with this prompt: “Identify recurring themes and pain points mentioned by multiple users. List themes with direct quotes and frequency.”
• Review the output skeptically
• Verify each theme by re-reading relevant sections myself
• Discard themes that feel AI-generated vs. human-real

Time saved: Maybe 1-2 hours on synthesis. But only if you still do the verification work.

3. Survey Data Analysis (With Major Caveats)

What I use: Google Sheets + ChatGPT for interpretation

Why it works: For quantitative data, AI can spot correlations and trends in large datasets faster than manual review.

The catch: AI is terrible at understanding why correlations exist. It’ll find statistical relationships that mean nothing in context.

My workflow:

• Export survey data to sheets
• Ask AI to identify significant patterns: “What correlations exist between user role and feature usage?”
• Manually investigate interesting patterns
• Interview actual users to understand why patterns exist

Time saved: Variable. Sometimes significant, sometimes it just creates more questions.

What AI Is Terrible At (Everything That Matters)

Now for the uncomfortable part: AI for UX research process fails at basically all the important stuff.

Understanding Context

AI can tell you what users said. It can’t tell you what they meant, what they didn’t say, or why something matters.

When a user says “this feature is confusing,” AI might categorize it as a usability issue. But if you were in the room, you’d notice they said it while laughing, after successfully using the feature three times. Context matters.

AI doesn’t have context. It has words.

Asking Follow-Up Questions

Good research is 90% follow-up questions. “Why?” “Can you show me?” “What were you trying to do?” “How did that make you feel?”

AI can’t do this. It can’t read body language, sense hesitation, or know when someone’s giving you a rehearsed answer vs. their real frustration.

You know what tool is great for asking follow-up questions? Your brain. Used in person. With an actual human.

Recognizing What’s Important

I did research for a SaaS dashboard redesign last year. In every interview, users mentioned they “checked the dashboard daily.” AI flagged this as high engagement.

But when I asked follow-up questions, they were checking it because the email alerts were broken. They hated the dashboard. They only used it because they had to.

AI saw “daily usage” and categorized it as positive. I saw the frustrated tone and understood it was a workaround for a broken feature.

Context. It matters.

Synthesis vs. Summary

AI can summarize. It’s great at summarizing. But synthesis — finding the deeper patterns, understanding what problems are worth solving, connecting user behavior to business goals — that’s human work.

Using AI for UX research to summarize is fine. Using it to synthesize is how you end up designing for patterns that don’t actually matter.

My Actual Workflow (Step by Step)

Here’s how I use AI tools for UX research in real projects without lying to myself about what’s happening:

Phase 1: Planning (No AI)

Research questions, participant criteria, interview guides — this is all human thinking. AI can’t help here because it doesn’t understand your specific product, users, or business goals.

If you’re using AI to write your research plan, you’re not doing research. You’re generating research theater.

Phase 2: Interviews (No AI, Obviously)

Talk to actual humans. Ask questions. Listen more than you talk. Take notes the old-fashioned way because the act of writing helps you process information.

I know some people record and transcribe automatically. I still take handwritten notes during calls because it forces me to synthesize in real-time.

Phase 3: Transcription (AI Shines Here)

Upload recordings to Otter or Descript. Get transcripts. This is where AI saves you hours of busywork.

Review transcripts while listening to key sections. Mark interesting quotes. Export everything to your synthesis document.

Phase 4: Initial Theme Identification (AI as Assistant)

Feed transcripts into Claude with a specific prompt:

“Review these 5 user interview transcripts. Identify recurring themes mentioned by at least 3 participants. For each theme, provide direct quotes and note which participants mentioned it. Flag any contradictions or disagreements between participants.”

Review the output. Verify themes by checking original transcripts. Discard themes that feel AI-invented vs. user-real.

This takes the first pass from 3 hours to 30 minutes. But you still need to do the verification work.

Phase 5: Deep Synthesis (Mostly Human, AI for Specific Tasks)

This is where you connect patterns to product decisions. Where you figure out what’s actually worth fixing. Where you understand not just what users said, but why it matters.

AI can help with specific tasks:

• “Pull all quotes related to onboarding confusion”
• “Find instances where users mentioned competitor products”
• “Identify all pain points related to mobile usage”

But the actual thinking about what this means for your product design? That’s human work.

Phase 6: Reporting (AI for Formatting, Not Content)

AI can help format findings, create summary slides, organize quotes. It cannot write your insights or recommendations.

If your research report could have been written by AI, you didn’t do research — you did keyword extraction.

The Tools I Actually Use (And Why)

For transcription:

• Otter.ai (free tier works fine for most projects)
• Descript (better for video interviews)

For synthesis:

• Claude (better at understanding context than ChatGPT)
• Google Docs (for actual synthesis work)
• Notion (for organizing findings)

For surveys:

• Typeform or Google Forms (basic, but they work)
• ChatGPT (only for analyzing quantitative data patterns)

For usability testing:

• Loom (for remote testing)
• Your eyes and brain (for analysis)

The Prompts That Actually Work

Generic prompts get you generic results. Here are prompts I actually use:

For Initial Theme Identification:

Review these interview transcripts. Identify:
1. Pain points mentioned by 3+ participants
2. Workflows that users described differently
3. Features users mentioned wanting vs. using
4. Contradictions between what users say and do

For each item, provide direct quotes and participant IDs.

For Quote Extraction:

Pull all quotes related to [specific topic]. Include:
- The full quote
- Which participant said it
- The context (what question prompted it)

Organize by sentiment: frustrated, neutral, positive.

For Pattern Recognition:

Compare these 5 participants' workflows for [specific task].
Identify:
- Steps all participants have in common
- Steps that vary significantly
- Points where participants got confused
- Workarounds participants created

Do not interpret or synthesize. Just map the patterns.

The key is being specific about what you want AI to do (pattern matching) vs. what you’ll do yourself (interpretation).

What Not to Do (I’ve Tried It, It Doesn’t Work)

Don’t use AI to write research questions. They’ll be generic and miss your specific context.

Don’t use AI to generate personas. You’ll get demographics and made-up pain points that sound plausible but aren’t real.

Don’t use AI to prioritize findings. It doesn’t understand your business constraints, technical limitations, or strategic goals.

Don’t use AI to write recommendations. This is where your UX UI design expertise actually matters. AI can’t make these calls.

Don’t use AI-generated insights without verification. Always check AI findings against original transcripts and your own observations.

The Model Collapse Problem for UX Research

Here’s something that’s going to get worse: AI model collapse is affecting user research too.

AI research tools are increasingly trained on AI-generated research reports. Which means they’re learning to produce research that looks like research, not research that’s actually useful.

You’ll see this in:

• Personas that hit all the demographic checkboxes but feel generic
• Insights that sound smart but don’t lead to actionable design changes
• Recommendations that could apply to any product in your category

The solution? Keep doing actual research with actual humans. Use AI for busywork, not for thinking.

Time Savings (The Honest Math)

Here’s what using AI for UX research actually saves me on a typical 5-interview research project:

Transcription: 10 hours → 1 hour (saves 9 hours)
Initial theme identification: 3 hours → 1 hour (saves 2 hours)
Quote extraction: 2 hours → 30 minutes (saves 1.5 hours)
_____
Total saved: About 12-13 hours per project.

What I don’t save time on:

• Planning research (because AI doesn’t know your product or users)
• Conducting interviews (unless you want to interview a chatbot, which, no)
• Deep synthesis (the part where you actually think about what it all means)
• Making recommendations (AI will suggest “improve the UX” — thanks, very helpful)
• Translating findings to design decisions (this is literally the job)

Those still take the same amount of time. Because they require human judgment, not pattern matching.

When to Use AI vs. When to Think

Good rule of thumb: if the task requires understanding context, making judgments, or connecting patterns to decisions, don’t use AI.

If the task is extracting information, identifying patterns, or organizing data, AI can help.

AI is good at:

• Transcription
• Pattern recognition
• Data organization
• Quote extraction
• Basic categorization

AI is bad at:

• Understanding why patterns matter
• Making strategic recommendations
• Prioritizing findings
• Connecting research to business goals
• Recognizing what’s important vs. what’s just frequent

The sweet spot is using AI to handle busywork so you can spend more time on the thinking that actually matters.

Why This Matters More Than You Think

There’s a growing trend of teams “doing research” by feeding AI a bunch of data and asking it to generate insights.

This isn’t research. It’s cargo cult research. Going through the motions without understanding what research is actually for.

AI can’t do that. It can help you with the grunt work so you have more time to actually think. But it can’t replace thinking.

If your research process could run without you, you’re not doing research. You’re generating reports.

The Practical Recommendation

Start small. Use AI for transcription. That’s the obvious win with basically no downside.

Then try using it for initial theme identification, but always verify its findings manually.

Don’t use AI for synthesis, recommendations, or strategic decisions until you’re extremely comfortable with its limitations.

And always remember: AI is a tool for doing research faster, not for skipping research entirely.

The goal isn’t to use AI as much as possible. It’s to use AI where it helps, and use your brain where it matters.

Because the hard part of UX research isn’t transcribing interviews or organizing quotes. It’s understanding what users actually need, why it matters, and what to do about it.

That’s still human work. And it’s not getting automated anytime soon.