How to do thematic analysis with AI: a practical guide for 2026

Analysing 40 interviews using Braun and Clarke's six-phase thematic analysis framework, done properly, takes somewhere between six and twelve weeks. Researchers familiar with the process know this is not an exaggeration. You read and re-read transcripts, generate initial codes, search for patterns, review candidate themes, refine their definitions, and write it all up. Each step requires sustained attention and genuine intellectual engagement with the data.

AI changes that timeline substantially. With the right approach, the initial coding pass that might take three weeks of concentrated work can be done in hours. But that speed comes with an important caveat: AI accelerates the mechanical work, not the interpretive work. What a theme means, why it matters, how themes relate to each other and to your research question, those decisions remain yours. They have to.

This guide is for researchers, consultants, HR professionals, and market researchers who already understand thematic analysis and want a clear-eyed account of what AI can genuinely do in this process, what it cannot, where the common approaches go wrong, and what a rigorous AI-assisted workflow actually looks like. It is not a beginner's introduction to thematic analysis; for that, see our complete guide to thematic analysis and demystifying thematic analysis.

What thematic analysis actually involves

Braun and Clarke's foundational 2006 paper in Qualitative Research in Psychology described six phases: familiarising yourself with the data, generating initial codes, searching for themes, reviewing themes, defining and naming themes, and producing the report. This framework has since become the most widely cited method for qualitative analysis across disciplines, partly because it is flexible enough to work across epistemological positions and research contexts.

What makes thematic analysis demanding is not primarily the reading. It is the constant movement between the data and the emerging conceptual structure. A code that seemed clear in the first five interviews starts to fragment into two distinct ideas by interview fifteen. A theme you thought was central turns out to be less important than a pattern you initially treated as background noise. Good thematic analysis is iterative and recursive, not linear.

The mechanical labour, however, is genuinely mechanical. Reading through transcripts and marking passages that seem relevant to a particular code does not require interpretive skill. It requires attention and consistency. This is where AI can help.

What AI can and cannot do in thematic analysis

What AI does well

Reading at scale. A researcher can hold only so much in working memory at once. AI can process an entire corpus of 50 or 100 interviews in a single pass, applying the same emerging category definitions consistently across all documents. This is harder for humans, particularly when analysis extends over several weeks and your understanding of the codes evolves.

Surfacing patterns across a large corpus. When you have 80 interviews and a theme appears in six of them, a human analyst working through data linearly may not notice the pattern until much later. Systematic AI processing can flag the co-occurrence of codes early, before interpretive commitments have hardened.

Consistent application of categories. Once a category has been defined, AI applies it without the fatigue, distraction, or unconscious drift that affects human coders working across long coding sessions. This does not mean AI coding is infallible. It means it is systematically consistent in a way that is difficult for humans to sustain.

Initial code generation, bottom-up. Given clean source data, AI can generate an initial set of candidate codes that a researcher can then work with. These codes are a starting point, not a conclusion. They often capture obvious patterns well and miss subtle or interpretive ones, but they reduce the time before a researcher has something concrete to engage with.

Where AI cannot substitute for researcher judgement

Deciding what is theoretically interesting. AI can surface patterns in data. It cannot tell you which patterns matter for your research question, your theoretical framing, or your audience. The move from "this comes up frequently" to "this is significant and here is why" is an interpretive act. No current AI system can make that move on your behalf.

Understanding the context of your data. The meaning of a phrase depends on who said it, in what context, in response to what question, and in the light of what else they said in the same interview. AI lacks the accumulated contextual understanding a researcher builds through immersion in a project. This is particularly important in organisational research, where what people say often reflects institutional pressures and political dynamics that are not visible in the transcript text itself.

Replacing interpretive judgement in theme development. Themes are not just buckets of similarly coded data. A well-developed theme is a claim about what a pattern means. The name, the definition, the boundaries, the relationship to other themes: all of these involve interpretation that must be exercised by the researcher.

Three common approaches that fail

Before describing what a good AI-assisted workflow looks like, it is worth being explicit about what does not work, because all three of these approaches are in common use.

1. Paste everything into ChatGPT and ask for themes

This is the most common approach and the most problematic. You export your transcripts, paste them into a chat interface, and ask the model to identify the main themes.

The problems are well-documented. Context window limitations mean large datasets get truncated or summarised into summaries of summaries, with significant information loss. Hallucinations are common at scale: the model produces quotes that do not exist in the source data, or attributes statements to patterns they do not actually support. There is no audit trail. You cannot trace from a theme back to the specific passages that generated it. And if you run the analysis again tomorrow with a slightly different prompt, you may get different themes. Practitioners have found that this inconsistency is a fundamental characteristic of how these models work, not a problem that better prompting can reliably address.

For a detailed analysis of why this approach fails for rigorous work, see can ChatGPT analyse qualitative data.

2. Use AI to apply pre-defined categories

Some researchers use AI in a more structured way: they define a set of categories in advance and ask the AI to code the data according to those categories. This avoids some hallucination problems and creates more reproducible results.

The limitation is that it defeats one of the main purposes of thematic analysis. Braun and Clarke's framework is designed to produce emergent themes that arise from the data, not to test a framework you already have. If you know what categories you are looking for, you may find them, but you will miss the unexpected patterns that often carry the most analytical value. This approach is closer to a deductive content analysis than to thematic analysis, and should be labelled accordingly.

3. Rely on keyword or sentiment tools

Keyword frequency and sentiment analysis are not thematic analysis, and they should not be presented as such. They miss context, nuance, and interpretive meaning. A negative sentiment score does not tell you what the person was actually expressing or why. A high frequency keyword may appear in radically different contexts across different interviews. These tools have their uses, but they do not substitute for the interpretive work that thematic analysis involves.

What a rigorous AI-assisted thematic analysis workflow looks like

The following workflow reflects both established qualitative methodology and what purpose-built tools can actually support. It differs from informal AI use in that it maintains a clear audit trail, keeps the researcher's interpretive role central, and produces outputs that are defensible in academic or professional contexts.

Step 1: Prepare your data

Thematic analysis begins before you open any analysis tool. Each interview or document should be a discrete file. Transcripts should be clean and readable: automated transcription is fast, but it introduces errors that can mislead pattern recognition, particularly for non-standard terminology or proper nouns. Decide at this stage whether you will include interviewer questions and prompts, or only respondent text. Both approaches are defensible; the important thing is to be consistent and to document your choice.

If your dataset includes heterogeneous document types (interviews alongside open survey responses, for example), think carefully about whether they belong in the same analysis or whether different data types should be handled separately before integration.

Step 2: Import with metadata

Metadata is what enables you to move beyond identifying themes to understanding their distribution. When you import your data, attach relevant metadata to each document: interview date, respondent segment, product line, geography, role, or whatever variables are meaningful for your research question.

This matters because a theme that appears in 80% of your interviews is analytically different from one that appears in 80% of interviews from one particular segment and rarely elsewhere. Without metadata, you cannot make that distinction. With it, you can cross-cut your thematic findings by segment to surface patterns that would be invisible in an undifferentiated corpus. Skimle's approach to discovering themes using metadata variables is designed specifically for this kind of structured interrogation.

Step 3: Let AI generate initial codes bottom-up

The key word here is bottom-up. Do not start with a predefined coding frame and ask AI to apply it. Instead, allow the AI to process each document and generate candidate codes that emerge from the content of the data itself. This preserves the inductive logic of Braun and Clarke's framework.

A well-designed system will process each document individually, generate codes at the passage level, and then aggregate those passage-level codes into candidate categories across the corpus. This is meaningfully different from asking a chatbot to summarise the overall themes: it operates at a more granular level, and each code retains its connection to the specific passage and document that generated it.

The output of this step should be a structured representation: a set of candidate codes and categories, each linked to the specific passages in the source documents that generated them. Understanding why this structured approach outperforms retrieval-based methods is explained in why RAG-to-riches does not work.

Step 4: Review, edit, merge, split, and relabel

This is where the researcher's interpretive work begins in earnest. The AI-generated codes are a starting point, not a conclusion.

Review the candidate categories. Some will need to be merged because they represent the same underlying idea expressed differently. Others will need to be split because what looked like a unified category actually contains two distinct phenomena. Many will need to be relabelled because the AI's language does not quite capture what the pattern is actually about.

Crucially, read the source passages as you do this. Do not rely on the AI's characterisation of what a passage means. Open the original document, read the quote in context, and form your own view. The categories that survive this review process will be grounded in your reading of the data, not just in the AI's processing of it.

This step corresponds to phases 3 and 4 of Braun and Clarke's framework: searching for themes and reviewing themes.

Step 5: Cross-cut by metadata to find distribution patterns

Once you have a provisional set of themes, use your metadata to interrogate their distribution. Which themes appear disproportionately in one segment? Which themes seem to be universal across the corpus? Are there themes that are prominent early in your data collection period but less so later, or vice versa?

These distribution patterns often contain analytically important information that flat theme lists obscure. They can also reveal whether your emerging themes are genuinely generalisable across your sample or are specific to particular sub-populations. For more on this, see our guide to discovering themes using metadata variables.

Step 6: Verify by reading source quotes for each theme

Before you write up, spend time with the source data for each theme. Go back to the actual passages. Do they support the characterisation you have given the theme? Are there passages you have assigned to a theme that, on reflection, belong elsewhere? Are there passages you have overlooked that are clearly relevant?

This verification step is not optional. It is what transforms an AI-generated draft into a researcher-endorsed analysis. It is also your insurance against hallucination: if a quoted passage does not exist in the source document, you need to know before you publish, not after.

Tools that link every code back to its source passage, and flag verification failures during processing, make this step substantially less onerous. The principle of two-way transparency (being able to trace from any theme to its supporting data, and from any document to the themes it contributes to) is what makes this kind of verification practical at scale.

Step 7: Write up with transparent methodological disclosure

Your write-up should describe not just what themes you found but how you found them. This includes the role AI played. See the section below on methodological disclosure.

Methodological disclosure: how to cite AI-assisted analysis

The question of how to disclose AI assistance in qualitative research has moved from a niche concern to a mainstream one with some speed. Most leading journals now have explicit policies, and the trend is towards requiring disclosure rather than leaving it to researcher discretion.

A recurring concern in qualitative methodology literature is that AI tools can create distance between the researcher and the data, a distance that risks undermining the reflexivity that underpins rigorous qualitative work. The implication is that AI use should be disclosed and that the researcher should be able to demonstrate continued engagement with the source material.

In practice, methodological disclosure for AI-assisted thematic analysis should address:

What tool was used, including the version where relevant
At which stages AI assistance was applied (initial coding, categorisation, both)
How the AI-generated outputs were reviewed and verified by the researcher
What manual interventions were made (merging, splitting, relabelling categories)
How source quotes were verified against original documents

Some journals, particularly in health research and management, now include specific checklist items for AI disclosure in their submission guidelines. If you are preparing work for peer review, check the journal's author guidelines before submission. The equator network reporting guidelines are beginning to incorporate AI disclosure requirements for qualitative research.

For consulting and professional contexts, methodological transparency serves a similar function. A client who asks how you reached a conclusion should be able to see the analytical process, not just the output. This is particularly important for sensitive work where stakeholder representations are at stake.

A practical approach for academic methods sections:

"Initial coding was conducted using [tool name], which processed each transcript individually and generated candidate codes at the passage level. All AI-generated codes were reviewed by [researcher name(s)], who merged, split, and relabelled categories based on close reading of source passages. Final themes were verified by manual review of all contributing quotes in their original context. The AI-generated initial structure is available as supplementary material."

This level of disclosure is defensible to reviewers, transparent to readers, and an accurate description of what a well-executed AI-assisted analysis actually involves.

For a broader discussion of how AI disclosure fits into the academic research context, see our guide on AI in qualitative research for academic researchers.

How Skimle approaches AI-assisted thematic analysis

Skimle was built specifically to support rigorous thematic analysis, not to replace it. The architecture reflects what proper qualitative methodology requires.

Each document is processed individually and coded at the passage level, producing a structured representation of what each interview or document says about each emerging category. This structure is stable. It does not change each time you query it, which means your analysis is reproducible and consistent in a way that chatbot-based approaches are not.

The two-way transparency principle is central to the design. From any theme, you can navigate directly to the specific passages that generated and support it. From any document, you can see every theme it contributes to. This makes the verification step described above practical rather than theoretical: you are not reading through 100 transcripts manually, but you are reading the specific passages that matter, in context, with a clear view of how they connect to the overall analysis.

Manual editing is a first-class feature, not an afterthought. You can rename, merge, split, and reclassify codes and categories. You can add codes manually for passages the AI missed. The AI-generated structure is explicitly a starting point for researcher interpretation, not a finished product.

For researchers who also work with traditional qualitative analysis software, Skimle supports REFI-QDA export for combining AI analysis with manual coding workflows, allowing you to move between tools without losing the work done in either.

You can read more about how Skimle works in what Skimle does.

AI as a collaborator in thematic analysis

The framing that works best for AI-assisted thematic analysis is neither "AI does the analysis" nor "AI is just a time-saving shortcut." It is closer to having a research assistant who can read everything, never tires, applies categories consistently, and produces a structured draft that you then work with.

A good research assistant does not replace the researcher's judgement. They extend the researcher's capacity to engage with a large dataset. They take on the mechanical work so the researcher can focus on the interpretive work. That is what well-designed AI does in thematic analysis.

The interpretive decisions (what themes mean, how they relate, what the analysis ultimately claims) remain with the researcher. They have to: those decisions require contextual knowledge, theoretical sophistication, and epistemic accountability that current AI systems cannot provide and, frankly, should not be asked to.

What has changed is the scale and speed at which a researcher can work. Fifty interviews no longer means three months of mechanical coding before the interpretive work can begin. That shift is significant. It changes what is feasible within a typical research timeline, and it changes what sample sizes are practically achievable for qualitative studies. For more on this, see our discussion of qualitative research sample size.

The epistemology of thematic analysis has not changed. The labour requirements have.

Ready to try a systematic AI-assisted approach to thematic analysis? Try Skimle for free and experience rigorous analysis with full two-way transparency from every theme back to the source data.

Want to go deeper on qualitative analysis methods? Read our guides on thematic analysis methodology, demystifying thematic analysis, and how to analyse interview transcripts.

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