To code qualitative interviews with AI, transcribe the recordings, then use a tool that systematically reads every passage and assigns codes tied to a verbatim quote, rather than a general chatbot that only summarises. Tools like Skimle code each transcript into a reviewable category structure and export to REFI-QDA for use in NVivo or MAXQDA.
Coding is the part of qualitative research where most of the time goes and most of the rigour is won or lost. It is also the part that people most often hope AI will take off their hands. The hope is reasonable, but the way it is usually attempted, by pasting transcripts into a chatbot and asking for the themes, produces something that looks like coding without actually being it. This guide explains what coding qualitative interviews really involves, why general-purpose AI struggles with it, how purpose-built AI coding works instead, and how to move the results into a classic environment like NVivo when you need to.
What does it mean to code a qualitative interview?
Coding is the act of attaching a short label to a specific segment of data so that meaning can be retrieved, compared, and counted later. A code is descriptive and close to the data. A theme is interpretive and sits above a set of codes that share a meaning. If a participant says "I kept waiting for someone to make a decision and it never came," a code might be "decision paralysis"; the theme that later gathers it up might be "unclear ownership."
Two features make coding different from summarising:
- Every code stays attached to its evidence. You can always click back from a code to the exact words that justified it.
- Codes are applied consistently across the whole dataset. The same idea gets the same label in interview 3 and interview 30, so you can compare and quantify.
Lose either property and you no longer have coded data. You have an impression. For the full picture of the paradigms involved, see our guide to how to code qualitative data and the deeper reference on qualitative coding.
Why can't ChatGPT code qualitative interviews properly?
General-purpose chatbots are built to generate fluent text, not to attach stable labels to fixed spans of source material. That difference matters more than it first appears. When you ask ChatGPT to "find the themes in these interviews," it produces a plausible summary, but it is not coding in the methodological sense, and several studies have documented why.
- It is inconsistent. Run the same transcript twice and you get different codes. Ana Canhoto, a marketing professor, found that ChatGPT is not consistent in its reading of the same interview data, which is fatal for a method that depends on applying the same label the same way every time.
- It approximates rather than analyses. Philipp Mayring, testing both ChatGPT 3.5 and 4 for content analysis, found the models produced at most rough approximations of a proper coded solution.
- It hallucinates evidence. Morgan (2023), writing in the International Journal of Qualitative Methods, reported that hallucination happens and that ChatGPT was less successful than a human analyst, inventing quotes or attributing them to the wrong speaker. When a "quote" does not exist in the transcript, the coding built on it is worthless.
- It forgets. A long interview set overruns the context window, so the model silently drops earlier transcripts and codes only what it can still see. You cannot know which material it ignored.
The reputational risk of trusting fluent-but-unfounded output is not hypothetical. In 2025, Deloitte was required to partially refund the Australian government after a report produced with AI contained fabricated references. For a fuller treatment of the failure modes, see hallucinations, context and the black box. A recent review by Nguyen and Welch (2025) in Organisational Research Methods reaches a similar conclusion: general LLMs can support qualitative work but should not be handed the analytical decisions unsupervised.
The short version: chatbots do not code text. They talk about it.
How is systematic AI coding different?
Purpose-built AI coding tools are engineered around the two properties that define coding, traceability and consistency. Instead of one sweeping prompt over a pile of transcripts, they break the job into hundreds of small, checked operations.
A tool like Skimle reads your corpus roughly a page at a time, extracts candidate insights with the exact supporting quote, verifies that each quote appears verbatim in the source, and only then organises the insights into a category structure. Because the pipeline is built from many atomic passes rather than a single generation step, it does not run out of context on a large interview set, and because every insight carries its verified quote, nothing is asserted that the transcript does not support. You can read how the pipeline is constructed in the docs on how analysis works and why a structured approach beats dumping everything into embeddings in why RAG-to-riches does not work.
The division of labour is the point: the AI does the systematic first pass across the whole dataset, and the researcher interprets, challenges, and refines what it surfaces. That is the design principle behind two-way transparency.
How to code qualitative interviews with AI in 5 steps
Step 1: Transcribe the recordings
You cannot code what is not text. If your interviews are audio or video, start by transcribing them. Rather than juggling a separate transcription service, you can add recordings directly and let the tool transcribe them for you: Skimle's built-in transcription handles audio and video across more than 100 languages, identifies speakers, and keeps the transcript in the same secure environment as the analysis, so there is no separate import or split billing. For a comparison of standalone options, see the best AI transcription tools for research.
Step 2: Choose inductive or deductive coding
Decide whether your codes should emerge from the data or be applied from an existing framework.
- Inductive (codes emerge from the interviews) suits exploratory work. In Skimle this maps to inductive analysis or the zero-prompt automatic thematic analysis.
- Deductive (you bring the codebook) suits confirmatory or applied work. This maps to predefined categories.
Most real projects are abductive, moving between the two. Our guide on inductive, deductive and abductive coding explains when to use each.
Step 3: Let the AI code every transcript
The tool now reads each transcript and produces coded insights, each backed by a verbatim quote and assigned to the most relevant category. This is the step that used to take weeks of manual work. Initial coding of 20 to 30 interviews takes minutes rather than weeks, which is the difference manual coders feel most acutely.
Step 4: Review, refine, and interpret
This is where your judgement does the important work. Read the categories, check quotes against their source, merge or split codes, rename anything the AI labelled clumsily, and add codes it missed. Because every insight links to its quote, verification is a click, not a re-read. See working with insights for the mechanics.
Step 5: Quantify and compare by subgroup
Once coded consistently, your interviews can be counted. While not essential for some approaches, in other cases this can yield interesting insights. Tag each transcript with participant attributes (e.g., role, tenure, segment, site) and you can see which themes concentrate in which groups. Skimle's metadata analysis flags where subgroup differences are statistically meaningful, so "senior staff raised trust more than juniors" becomes a defensible finding rather than a hunch.
Can you take AI coding back into NVivo or MAXQDA?
Yes, and this matters for anyone who codes in a traditional environment or needs to hand data to a co-analyst who does. Coding your interviews with AI does not lock you in.
Skimle exports to REFI-QDA, the open interchange standard (the .qdpx format) recognised by NVivo, MAXQDA, ATLAS.ti, Dedoose, and QDA Miner. Your documents, your full code hierarchy, and every coded segment mapped to its position in the source text move across intact. A common hybrid workflow is to let AI do the systematic first pass, export to REFI-QDA, and finish the interpretive coding by hand in your usual tool. We describe this in detail in combining Skimle AI analysis with manual REFI-QDA workflows.
ChatGPT vs dedicated AI coding vs manual QDAS
| Dimension | General chatbot (ChatGPT) | Dedicated AI coding (Skimle) | Manual QDAS (NVivo, MAXQDA) |
|---|---|---|---|
| What it actually does | Summarises text | Codes every passage with verified quotes | You code by hand |
| Traceability to source | Weak, quotes can be invented | Every insight links to a verbatim quote | Full, you place each code |
| Consistency across dataset | Low, varies run to run | High, systematic passes | High, but depends on coder fatigue |
| Handles 50+ interviews | Runs out of context | Yes, atomic passes | Yes, but very slow |
| Speed | Minutes, but unreliable | Minutes, reviewable | Weeks |
| Interoperability | None | REFI-QDA export | Native |
The takeaway is not that chatbots are useless, but that they belong in ideation and drafting, not in the coding step where evidence and consistency decide whether your findings hold up. For the wider tool landscape, see our full comparison of qualitative analysis tools.
Frequently asked questions
Can AI replace manual coding entirely?
No, and it should not. AI can do the systematic first pass across the whole dataset far faster than a human, but interpretation, deciding what a pattern means and whether it answers your research question, remains the researcher's job. The most defensible workflow keeps a human reviewing and refining the AI's output.
Is AI-assisted coding acceptable for academic publication?
Increasingly yes, provided you are transparent about the method and can show traceability from each finding to its source data. Report the tool, the approach (inductive or deductive), and how you reviewed the output. Reviewers care that quotes are verifiable and that a human made the analytical decisions.
Does AI coding work for focus groups and multi-speaker recordings?
Yes. The transcription step identifies speakers, and each utterance is coded like any other passage. Attribution in group settings is harder than in one-to-one interviews, so review speaker labels carefully. See analysing focus group transcripts.
What about sensitive interviews with personal data?
Remove identifiers before wider analysis. Skimle Anonymise detects names, roles, locations, organisations and dates and pseudonymises them consistently across the set, with an audit trail for ethics boards.
How many interviews do I need before coding is worthwhile?
Coding is worthwhile from the first interview, but patterns stabilise later. In their study in Field Methods, Guest, Bunce and Johnson (2006) found that saturation was largely reached within the first twelve interviews for a fairly homogeneous sample. See how many interviews are enough.
Ready to code your interviews with speed and full traceability? Try Skimle for free and see systematic AI-assisted coding where every insight links back to a verbatim quote, with REFI-QDA export when you want to continue in NVivo or MAXQDA.
Want to go deeper? If you work in academic research, see how Skimle fits your workflow, and read our guides on how to code qualitative data and whether ChatGPT can analyse qualitative data.
About the authors
Henri Schildt is a Professor of Strategy at Aalto University School of Business and co-founder of Skimle. He has published over a dozen peer-reviewed articles using qualitative methods, including work in Academy of Management Journal, Organisation Science, and Strategic Management Journal. His research focuses on organisational strategy, innovation, and qualitative methodology. Google Scholar profile
Olli Salo is a former Partner at McKinsey & Company where he spent 18 years helping clients understand the markets and themselves, develop winning strategies and improve their operating models. He has done over 1000 client interviews and published over 10 articles on McKinsey.com and beyond. LinkedIn profile
