UX research generates a mountain of qualitative data — interview recordings, open-ended survey answers, support tickets, usability notes — and the slow part has always been turning that mess into clear, defensible insight. AI does not change what good research is, but it can collapse the days you spend transcribing, coding, and clustering into hours, freeing you to spend more time on the parts only a human can do: deciding what to study and judging what the findings mean.
This guide covers four practical workflows — synthesising interview transcripts, auto-coding qualitative data, generating personas from real data, and building affinity maps — with the guardrails that keep AI-assisted research honest.
Synthesising interview transcripts
Start by getting clean transcripts. Modern speech-to-text models transcribe an hour of audio in a minute or two with speaker labels. Once you have text, an LLM can produce a first-pass summary of each interview: top goals, key frustrations, memorable quotes, and unexpected moments.
The discipline that makes this trustworthy is grounding. Ask the model to attach the verbatim quote behind every claim, like “struggled to find the export button (00:14:32)”. If it cannot quote the source, the claim is suspect. Never accept a synthesis that paraphrases without evidence — that is exactly where hallucinated themes creep in.
Auto-coding qualitative data
Thematic coding — tagging each chunk of data with a theme — is the most tedious part of analysis and the most natural fit for AI. Give the model your codebook (or ask it to propose one), then have it apply codes across all transcripts at once. A useful prompt: “For each highlighted statement, assign one or more codes from this list, and quote the exact text. If nothing fits, mark it ‘uncoded’ rather than forcing a code.”
The key safeguard is the uncoded escape hatch. Without it, the model will stretch your existing codes to cover everything and you will miss the surprising signal that did not fit. Review the uncoded pile by hand — that is often where the most interesting findings live.
Generating personas from real data
AI personas earn their keep only when they summarise real segments rather than invent characters. Feed the model aggregated, de-identified data — behaviours, goals, contexts, and how frequently each pattern appears — and ask it to cluster participants into groups and describe each group.
A trustworthy persona traces every trait back to evidence. If the model writes “prefers mobile,” there should be data showing it. Strip out anything it cannot support, and resist the temptation to add a stock photo and a cute name that implies more certainty than the data warrants. The persona is a research artifact, not a marketing illustration.
Building affinity maps
Affinity mapping clusters individual observations into themes. Paste your coded observations and ask the model to group similar items and name each cluster. This produces a fast draft you can then rearrange — AI is good at the first sort, humans are better at the judgement calls about where a borderline item belongs.
Treat the AI output as a starting layout, not the final map. Move things, split clusters that are too broad, and merge ones that overlap. The goal is to skip the blank-canvas stage, not to outsource the thinking. Used this way, AI turns UX synthesis from a week of sticky notes into an afternoon of refinement — without ever letting the researcher off the hook for what the data actually says.