AI UX Research Synthesis: What We Let the Model Do and What a Senior Researcher Checks
Every team that runs user interviews now feeds transcripts to a model. Ten interviews go in, a tidy list of themes comes out, and the deck is ready by lunch. The speed is real. So is the risk. A model will happily give you a quote nobody said, credit the interviewer's words to a participant, or turn one frustrated person into "users feel."
We use AI in research synthesis on every engagement. We also treat its output as a first draft that has to earn its way into a recommendation. This post lays out exactly where that line sits in our studio, why we drew it there, and what we refuse to do.
Our rule: the model sorts, a senior researcher decides
The short version: AI handles volume, and a senior person owns meaning. Anything that changes what a team builds has to trace back to a real person saying a real thing, and a senior researcher has to have looked at it.
That rule shapes the whole workflow. We let models do the work that scales with the number of hours of tape. We keep humans on the work that scales with consequences.
What the model does in our synthesis
Transcription and cleanup. Speaker labels, timestamps, filler removal. This is mechanical work, and modern tools do it well. We still spot check speaker labels, because a swapped label poisons everything downstream.
First pass coding. We ask the model to tag each transcript against the research questions we wrote before the first interview. Tags come back with participant ID, timestamp, and the verbatim line, in that order. Asking for the source first and the quote last makes fabrication easier to catch, because a quote with no traceable location stands out immediately.
Clustering and draft themes. Once tags exist, the model groups them into candidate themes and counts how many participants support each one. The count matters more than the wording. A theme backed by two of twelve people is a lead to watch. A theme backed by nine of twelve is a finding.
Contradiction hunting. This is one of the most useful jobs we give a model, and few teams ask for it. We prompt it to find every place where a participant contradicts a draft theme, or contradicts themselves. Models are good at agreeing with you. Asking them to argue with you is where they earn their keep.
Retrieval during the readout. When a stakeholder asks "did anyone mention pricing?", a searchable, tagged corpus answers in seconds. That keeps the conversation grounded in evidence and away from whoever remembers the loudest session.
What a senior researcher checks, every time
Every quote in the final report gets verified against the transcript. Every single one. If a quote supports a recommendation, a person has read it in context, at the timestamp, with the lines before and after. This is the step teams skip when they are tired, and it is the step that catches the worst errors.
Who said it. Models blur roles. The interviewer's leading question becomes the participant's opinion. A sales engineer sitting in on the call becomes "a user." We confirm the speaker and the role for every piece of evidence we cite.
Weight, beyond frequency. Three people mentioning a confusing label is a usability note. One compliance lead explaining that a missing audit trail blocks the purchase is a deal breaker. A model counts mentions. A senior researcher who knows regulated products knows which single sentence should reorder the roadmap.
What was not said. Silence is data, and models are weak at noticing it. If nobody in a fintech study mentioned fees, that could mean fees are clear, or it could mean participants never reached the screen where fees appear. Only someone who watched the sessions and knows the flow can tell which.
Leading questions and our own bias. We review our own interview guide against the findings. If a theme lines up suspiciously well with how we phrased a question, we downgrade it. A model has no reason to flag that it is echoing our wording back to us.
The calls we never hand to a model
Some decisions stay with people regardless of how good the tools get.
- Choosing the research questions. What we need to learn depends on the business bet, the risk, and the stage of the company. That is a judgment call made with the founding team.
- Recruiting criteria. Who we talk to decides what we hear. Getting this wrong makes every downstream step precise and useless.
- Turning findings into recommendations. A theme says what people experienced. A recommendation says what to build, what to cut, and what to test next. That requires product context, technical constraints, and taste.
- Anything involving sensitive data. Health details, financial records, and identity documents come up in research for regulated products. We strip and control that data before any of it touches a tool, and we decide case by case what is allowed in.
- Telling a team something they do not want to hear. Synthesis that only confirms the roadmap is a warning sign. Delivering an uncomfortable finding with enough evidence to hold up in the room is senior human work.
Our reject list
Practices we see in AI research synthesis that we will not use:
- Themes without receipts. A theme with no participant IDs or timestamps attached does not go in a report.
- "Users feel" summaries. Vague aggregate sentiment hides the n. We write "seven of ten participants" or we do not make the claim.
- Synthetic participants as evidence. Simulated users can help draft an interview guide or stress test a hypothesis. They never count as findings, because no real customer said anything.
- One prompt, one pass, ship it. A single summary prompt over all transcripts is the fastest way to get confident fiction. We work in stages, with checks between them.
- Letting the tool pick the quotes. Quotes in a readout are persuasion. A person chooses them, and chooses ones that represent the theme fairly, including the awkward ones.
Why we built it this way
Startups usually run research under pressure. There is a launch date, a board meeting, or a fundraising round, and the temptation is to let the fastest tool produce the final answer. The cost of a wrong finding in a complex product is high. A misread onboarding study in fintech can mean weeks building the wrong verification step. A misattributed quote in a health product can send a team after a problem patients do not have.
AI gives us back the hours that used to go into tagging and sorting. We spend those hours on the parts that carry the risk: verifying evidence, weighing it, and turning it into decisions a team can build on. That is why our engagements are staffed with senior people only. The model makes a senior researcher faster. It cannot make a junior one senior.
It also connects to how we work in general. We prefer working software over static mockups, and we prefer evidence over opinion. AI synthesis that cannot be traced back to a person fails the same test a pretty Figma file with no behavior fails: it looks finished, and nobody can check it.
A quick self check for your own team
If your team already uses AI to synthesize research, ask these questions about your last study:
- Can you trace every quote in the readout to a participant and a timestamp?
- Does each theme show how many participants support it?
- Did anyone look for evidence that contradicts the main themes?
- Was the interviewer's speech kept separate from participant speech?
- Did a senior person decide the recommendations, or did they come straight from the summary?
If any answer is no, your synthesis may be fast and wrong at the same time. The fix is a review step with a named owner, and it usually costs a few hours.
Work with us
If you are about to run discovery interviews, or you have a pile of transcripts and a summary you are not sure you trust, we can help. We run research and synthesis for complex products in fintech, health, insurance, AI, marketplaces, and B2B SaaS, with senior people on every call and AI doing the heavy sorting.
Book a discovery call or email ash@uxsignalstudio.com.