The task
You have a stack of interview transcripts, customer support chats, survey open-ends, or diary studies. You need to identify recurring themes — what users call “the thing” — and turn dozens of hours of conversation into a defensible set of codes, definitions, and representative quotes.
When AI is the right tool
- You have 10-100 transcripts and manual coding would take weeks.
- You already have a coding framework (open, axial, or deductive) you want applied consistently.
- You need a first pass so a human researcher can refine, not start from zero.
- Your team needs faster signal between research cycles to keep product moving.
When not to rely on AI alone
- Academic publications where methodological transparency is required.
- Sensitive topics (medical, legal, abuse) where misinterpreting a quote has consequences.
- Small samples (under 5 transcripts) — manual coding is faster and more accurate.
- Cultures or languages the model has limited training data for.
What to feed the AI
- The transcripts, with speaker labels intact and identifiers anonymized.
- The coding framework — predefined codes, or instructions to do open coding.
- Research question in one sentence (“what blocks first-time users from completing setup?”).
- Excerpts of how you have coded similar data before, as exemplars.
- Stop-list: codes that are too general to be useful (“user feedback”, “general comment”).
Copy-ready prompt
You are assisting a qualitative researcher with thematic coding.
Research question: {research_question}
Coding approach: {open_or_axial_or_deductive}
Predefined codes (if any): {predefined_codes}
Stop-list (codes to never use): {stop_list}
Exemplar coding from prior data: {exemplars}
Transcripts to code:
"""
{transcripts}
"""
Output:
1. A table of codes:
- Code name (2-4 words)
- 1-sentence definition
- 2 verbatim quotes that anchor the code (with speaker + line reference)
- Frequency across transcripts
2. A short axial section: which codes cluster into 3-5 higher-order themes.
3. A "boundary cases" list: 3-5 quotes that were hard to code, with your reasoning.
4. A flag list: any code where you are less than 70% confident.
Rules:
- Quote text must appear verbatim in the source. Do not paraphrase.
- Cite speaker and line number for every quote.
- If a quote does not fit any code, place it in "uncoded — needs human review".
- Do not invent codes that lack at least 2 supporting quotes.Recommended output structure
A code table with definitions, quotes, and frequencies; a themes block; a boundary-cases block; a low-confidence flag list. This mirrors how a researcher would document their codebook in NVivo, Dedoose, or Atlas.ti.
How to check the output
- Pick 10% of segments at random and double-code them yourself. Compute simple agreement.
- Verify every quote is actually in the transcript and attributed to the right speaker.
- Stress-test boundary cases with a colleague — disagreement is where insight lives.
- Sanity-check frequencies: if a code only appears once, it is an observation, not a theme.
Common mistakes
- Letting the model invent codes that sound good but lack supporting quotes.
- Accepting paraphrased quotes — they are unusable for reporting.
- Skipping inter-rater reliability because the model “sounded confident”.
- Coding at too coarse a level so all themes blur into “users want better UX”.
Next steps to keep improving
Build a project-specific codebook over multiple rounds. As exemplars accumulate, model agreement with human coders improves. Track agreement quarter over quarter; aim for >80% before publishing findings externally.
Practical depth notes
For AI Qualitative Coding: Code Interview Transcripts Like a Trained Researcher, the difference between a usable AI result and a generic one is the input packet. Give the model the audience, the current draft or raw material, the desired format, the decision you need to make, and two examples of what good and bad output look like. Ask it to preserve facts first, then improve structure or wording second.
After the first response, do a separate review pass. Look for missing constraints, invented details, weak calls to action, and language that sounds plausible but does not match the real situation. The best final output should be easy to use immediately: clear owner, clear next step, and no hidden assumption that someone else has to untangle. A stronger version of this workflow also defines the handoff. Decide who will use the output, what they should do next, and what information would make them reject it. If the deliverable is copy, test whether it has a single clear action. If it is analysis, test whether it separates observation from recommendation. If it is planning, test whether dates, owners, and tradeoffs are explicit enough for someone else to execute. One final check: compare the finished result against the original goal in a single sentence. If that sentence is hard to write, the output is probably polished but unfocused. Tighten the goal, remove decorative language, and rerun only the weak section instead of regenerating the entire piece.
FAQ
- Can AI replace qualitative researchers? No. It accelerates coding but cannot interpret context, irony, or what is unsaid.
- What about inter-coder reliability metrics? Compute Cohen’s kappa or simple percent agreement on a 10% sample.
- Which transcripts should I include in one prompt? Group by participant segment or interview wave to keep context coherent.
- How do I handle multilingual data? Code in the source language; translate only the quotes you cite in the report.
Related
For complementary patterns, see user feedback clustering prompts, the user feedback clustering AI workflow, and customer discovery questions AI.
Tags: #Data analysis #Workflow #Research