AI user research: what holds up, and what breaks
AI can run the interview now, not just transcribe it. That changes which questions are worth asking and which are not. Here is what survives contact with real users, where it fails, and how to run a study so the findings hold up when somebody pushes back.
Three things sold under one name
Analysis assistants take research you already ran and cluster it. Useful, but the weeks went into running the interviews, not sorting them afterwards.
AI moderation runs the interviews themselves. The model asks your question, hears the answer, and decides what to ask next. This is the part that changes the timeline.
Synthetic respondents generate the people too: a model predicting what a user like this would probably say. It is fast and it demos well. It is also a model quoting its training data back to you, and it cannot tell you about the bug your users hit last Tuesday. Know which of the three you are buying.
Where it holds up
Volume, obviously. A researcher manages about five interviews a day. Hundreds can run at once, which makes a question worth asking that was never worth three weeks of fieldwork.
Consistency is the quieter advantage. A moderator on their ninth call of the week leads a little, skips a little and gets tired. The AI asks the ninth participant what it asked the first.
People are often blunter with a machine. On money, on embarrassment, on why they really left, there is nobody to disappoint — and it shows in the transcripts.
And language stops being a budget line. The same study runs in 18 languages at once, so a market is not excluded because the agency quote was too high.
Where it breaks
A bad question gets asked perfectly, hundreds of times. A human moderator notices the question is not landing by the third call. An AI does not, so the cost of a vague brief shows up at the end, in the findings, rather than at the start.
It cannot watch. If what you need is someone trying to complete a task in your interface, that is usability testing and this is the wrong tool.
Recruitment still decides the quality of everything. Hundreds of interviews with the wrong people is a faster way to be wrong.
And it will not decide for you. A machine can rank themes by how often and how strongly they appeared. Which one matters for the decision in front of you is still judgement.
How to run one so the findings hold
Ask one question, not twelve. The studies that come back vague are nearly always the ones that tried to cover a roadmap in a single interview.
Say what would count as a complete answer before you start, then let the interview keep going until it has one. That is the difference between a questionnaire and an interview.
Interview close to the event. Someone who cancelled yesterday remembers the reason. Someone who cancelled in March has rehearsed a story about it.
Read three transcripts yourself, in full, before you read any summary. It takes fifteen minutes and it is how you catch a question that was heard differently from how you meant it.
Then open the evidence behind the top theme. If a finding cannot show you the passages that produced it, treat it as an opinion that has been through a model.
Where we fit
8xResearch runs the moderation part: hundreds of interviews at once, by voice or text, each one following up where an answer is thin, with every theme linked to the passages behind it. If what you need is a repository, an analysis assistant or a usability test, those are different tools and they will serve you better.
Try a study now