Gains:
- Ability to transform raw interview transcripts, open-ended survey responses and support tickets into themes and insights with artificial intelligence
- Ability to accelerate qualitative analysis steps such as affinity mapping and coding with artificial intelligence support and verify the results with the human eye
- Ability to weed out themes made up or exaggerated by artificial intelligence by linking them back to real quotes
The most tedious and most easily skipped phase of user research is synthesis: turning dozens of interviews, hundreds of survey responses, and support tickets (the user's request for help) into meaningful insights. This job takes days by hand; That's why teams often cut corners and decide "by feel." AI opens up just this bottleneck: it breaks down raw text into themes in minutes instead of hours. But synthesis is not a "produce summary" task; It is an evidence-based, auditable reasoning. This unit shows you how to preserve the authenticity of insight while making AI your synthesis partner.
What is insight, not summary
Let's clarify the term first. Insight is a meaningful pattern that emerges from the data and guides the design: "Users give up not at the checkout step, but when they first see the shipping fee." This is not a summary ("Users complain about payment"); nor is it an observation (“Participant 4 stopped when he saw the shipping fee”). Insight captures the common cause underlying many observations and opens the door to a design decision.
AI is naturally inclined to produce summaries. Your job is to pull it from summary to insight: “why?” and “what does this change in the design?” by embedding your questions into the prompt.
Qualitative analysis steps and the place of artificial intelligence
Qualitative (non-numerical, text/observation-based) analysis classically proceeds with these steps:
- Coding: Each meaningful piece in the text is given a label ("shipping cost concern", "lack of confidence").
- Affinity diagram (affinity mapping): Similar tags are grouped; groups become themes.
- Theme prioritization: Which theme is seen by how many users and with what intensity is determined.
- Insight and advice: Themes "what to do?" connects to sentences.
AI greatly speeds up steps 1 and 2: delivering initial coding and initial grouping in minutes. However, the meaning of the groups, the importance of the themes, and the accuracy of the suggestions require human verification. The AI gives you a “first draft map”; You correct the map according to reality.
Tip: Before you tell the AI to "extract themes", give it your coding scheme (list of tags). In this way, it works in everyone's common language and the outputs are comparable.
Chain of evidence: each theme links to a quote
The golden rule of synthesis: every theme is linked to at least one verbatim quotation. The quote is proof that the theme is real. Always include the citation requirement when asking the AI for a theme; Tell him not to produce a theme for which he cannot find a quote. This way you stop the hallucination at the source: a made-up theme is marked as "unfounded" because it has no counterpart in the source.
layer
example
source
observation
"Participant 4 abandoned the cart when he saw the shipping fee"
one-on-one recording
Code
"unexpected cost"
Labeling
theme
"Hidden costs break trust"
repeat in 5 participants
insight
"Shipping cost should be shown early"
I leave the theme
Suggestion
"Add shipping estimate to product page"
Design decision
This chain starts when a stakeholder asks “how do we know this?” When you ask, it allows you to go back from observation to quote.
three mini cases
Case 1 — 240 survey responses, 90 minutes. One team AI-coded 240 open-ended (free text) survey responses. The model suggested 11 themes; The team reviewed these in 90 minutes, narrowed them down to 7, combined 2 themes, and removed 2 as unsupported. A 2-day manual job was reduced to half a day and the chain of evidence was preserved.
Case 2 — Exaggerated theme caught. AI produced a strong theme of “users hate the app.” When the team looked at the quotes, they found that only 2 participants were harsh, while 18 were neutral. The theme was rewritten as "intense disappointment in a certain flow". Lesson: AI can exaggerate emotional intensity; count the frequency.
Case 3 — Missing minority opinion. While the model highlighted majority themes, it downplayed a critical accessibility issue raised by 3 participants as “unimportant.” The designer noticed this and preserved it as a separate find; It later became clear that this issue affected all screen reader users. Lesson: minority signal is not always "less important".
Copiable prompts
Your role: qualitative research analyst.Coding scheme (tags): <<list>>Task: Code the anonymized transcript below with these tags.Add [Participant
Transform the following coded data into a proximity diagram: group similar codes, name each group a theme. For each theme: count how many individual participants it appears in, include the 2 strongest quotes, and write a one-sentence "why" explanation. Mark themes with weak evidence separately.Data: <<codes>>
I want to test the following theme: "<<theme>>".List separately the quotes that SUPPORT and REFUSE this theme in the source text.At the end, evaluate how strongly the theme is supported (strong/medium/weak).Source: <<transcription>>
Turn this list of themes into insights and suggestions. For each theme:1) one-sentence insight (containing “because…”),2) a concrete design suggestion,3) what screen/flow this suggestion affects.Don't make up; If the basis is weak, write "more research is needed." Themes: <<list>>
Weak prompt / Strong prompt
Weak: "Analyze these conversations and tell me the results."
Result: A summary without sources, without frequency information, and open to exaggeration.
Strong: "Code this transcript with the labels I gave you; connect each theme to the quote from [Participant
Result: A high-frequency, evidence-based, auditable synthesis.
Difference: strong prompt prompts frequency + quote + weak anchor prompt; these prevent exaggeration and hallucination.
Common mistakes
- Not counting frequency. Making a decision without knowing whether there are 2 or 20 people behind the sentence "Users want X".
- Taking the intensity of emotion as it is. AI tends to highlight strong expressions; It can hide a neutral majority.
- Eliminating the minority signal. An issue that few people experience may be critical (e.g. accessibility).
- Not requiring citation. The quoteless theme is the unverifiable theme.
- Finish in one lap. Synthesis is iterative; Test the first output and sharpen the prompt.
In summary
Synthesis is the reasoning that turns raw data into insight that drives design. AI greatly speeds up mechanical steps like coding and proximity grouping, but the meaning and significance of the insight remains human. The golden rule is to connect each theme to a literal quote; count frequency, correct for emotional exaggeration, don't miss the minority signal. When you establish a chain of evidence that extends from observation to quote, from theme to suggestion, you achieve a research output that is both fast and defensible.
Application task
- Anonymize your 8-10 (or fictional) interview/survey responses.
- Write your own coding scheme (6-8 labels) and have it coded by artificial intelligence with the first prompt.
- Remove themes with the second prompt; Note how many participants appear for each theme.
- Test the strongest theme with the third prompt: compare supporting and refuting quotes.
- Write a theme in the form of insight + suggestion + affected screen and chart the chain of evidence.
checklist
- [ ] I gave my coding scheme to the AI in advance.
- [ ] I linked each theme to at least one verbatim quote.
- [ ] I counted how many individual participants appeared for each theme.
- [ ] I often corrected exaggerated emotional expressions.
- [ ] I made a conscious decision whether to eliminate minority signals.
- [ ] I linked the insights to the concrete design proposal and the affected screen.