What AI actually does well in user research
AI is genuinely transformative for the grunt work of qualitative research, and nearly useless for the judgment. It excels at four tasks: transcribing recorded interviews into searchable text, clustering raw notes into candidate themes, tagging quotes by topic or sentiment, and drafting a first-pass synthesis you can edit. Each of these used to consume hours per study. What AI does not do is decide what to study, design good non-leading questions, read the room, or judge whether a participant is representative. Treat AI as a fast research assistant that never gets tired — not as the researcher.
Dedicated platforms: Dovetail and UserTesting AI
Dovetail is a research repository first. Its AI features — auto-transcription, theme suggestions, sentiment tagging, and “ask your data” search — sit on top of a persistent store of every study, highlight, and tag your team has ever created. That persistence is the real value: insights from a study six months ago remain searchable and linkable, so you build institutional memory rather than one-off decks. UserTesting AI leans into moderated and unmoderated testing at scale, auto-summarising sessions, flagging friction moments in screen recordings, and surfacing sentiment shifts across many participants quickly. Both shine when you run continuous research across a team and need governance, sharing, and an audit trail more than raw model flexibility.
General LLMs: ChatGPT, Claude, and Gemini
For a researcher analysing one or a handful of interviews, a general model is often the pragmatic choice. Paste a cleaned transcript and ask for themes with supporting verbatim quotes, an affinity-style grouping, or a summary aimed at a specific stakeholder. Claude and Gemini handle long transcripts well thanks to large context windows; ChatGPT is strong for iterative back-and-forth synthesis. The cost is near-zero and the flexibility is total. The downside is everything the platforms provide: no persistent repository, no team collaboration, no automatic linkage between a quote and its source timestamp, and no governance — you are responsible for storing and structuring the output yourself.
How to choose
Pick by scale and persistence, not by features. A solo researcher or a small team running occasional studies gets most of the value from a general LLM plus a transcription tool, at a fraction of the cost. A research team running continuous studies, sharing findings across product squads, and needing to revisit old insights should invest in a repository tool like Dovetail or a testing platform like UserTesting. A common hybrid works well: use the platform as the system of record and a general LLM for ad-hoc deep synthesis on top.
Guardrails that matter regardless of tool
Whichever you choose, enforce three rules. Demand verbatim quotes for every theme so synthesis stays anchored to what people actually said. Review the outliers the model dropped — the single dissenting voice is often the most valuable signal in the room. And keep a human reading the raw sessions, because the strategic judgment about what a finding means for the product is precisely the part AI cannot do. Used this way, AI turns days of analysis into hours without quietly laundering bias into your conclusions.