Manual thematic analysis
A careful thematic analysis begins before a codebook exists. In the first phase, Familiarizing yourself with the data, the researcher reads transcripts, listens to audio when possible, records first impressions, and notices where the material feels surprising, emotionally charged, contradictory, or central to the research question. Familiarization begins interpretation because the researcher starts to understand the dataset as a whole rather than as disconnected quotations.
The second phase is Generating initial codes. Codes mark meaningful features of the data. They may be semantic, staying close to explicit wording, or more latent, pointing toward assumptions and meanings beneath the surface. If an interview participant says remote work gives them quiet but reduces opportunities to observe senior colleagues, possible codes might include "protected focus," "missing tacit learning," and "productive isolation." Those codes are not final themes. They are analytic handles that let the researcher return to evidence.
The third phase is Searching for themes. Here the researcher asks whether multiple codes and excerpts can be organized into candidate patterns. A theme might begin as a cluster around "autonomy with hidden developmental cost." At this stage, the wording is provisional. The researcher is testing whether a pattern is meaningful for the research question instead of grouping every similar topic into a bucket.
The fourth phase is Reviewing themes. Candidate themes are checked against the coded excerpts and the full dataset. Some themes collapse because they are too thin. Some split because one label is hiding two different analytic stories. Some merge because the boundary between them is artificial. This is where many weak analyses fail: a theme that sounds plausible in a slide title may not survive a return to the transcripts.
The fifth phase is Defining and naming themes. A theme name should do more than point to a topic. "Remote work" is a topic; "autonomy can mask weakened apprenticeship" is closer to an analytic theme because it says what pattern the researcher is claiming to see. The definition should state what the theme includes, what it excludes, how it relates to other themes, and why it matters for the study.
The sixth phase is Producing the report. Writing often reveals gaps in the analysis. If a paragraph cannot connect a theme to vivid evidence, the theme may need more review. If a quote is memorable but does not support the claim being made, it should not be used as decoration. A strong report makes the analytic path visible enough for readers to understand why the interpretation is credible.
The two methodological points AI writing often gets wrong
First, these six phases are not a linear process that is completed once. Reflexive thematic analysis is recursive. Researchers move back from candidate themes to codes, from report writing to data familiarization, and from theme definitions to the raw transcripts. A theme that looks strong early may become weaker when later interviews complicate it. A code that seemed minor may become important after a new case makes the pattern visible. The method expects this movement.
Second, in reflexive thematic analysis, themes are actively developed or constructed by the researcher. They do not simply emerge from the data like objects waiting to be collected. This matters because it rejects the idea that a tool can automatically discover final themes for you. The researcher brings theoretical commitments, research questions, interpretive judgment, and reflexive awareness to the work. Software can help organize the evidence, but it does not remove the researcher's role in constructing meaning.
Software-assisted thematic analysis
QDA software helps by giving the analysis a durable workspace. During familiarization, transcripts, audio, notes, and source metadata can sit together. During initial coding, the researcher can highlight passages, attach codes, write memos, and retrieve every excerpt under a code. During theme development, matrix views, filters, code groups, and memo links can help the researcher compare cases and notice where a candidate pattern is strong or weak.
Use software to make interpretive work traceable. When you rename a code, merge two categories, or discard a theme, the project should preserve enough context that you can explain the decision later. This is especially important in thesis work, team research, and client studies where another person may ask how you moved from transcripts to claims.
For researchers comparing tools, it helps to read a workflow-level guide such as traditional QDA workflow vs OpenVerbatim. Traditional tools often center manual coding operations, code trees, memos, and retrieval. OpenVerbatim starts from a review pipeline in which coding suggestions and theme suggestions remain tied to evidence until a reviewer confirms them. Both models still require methodological judgment.
AI-assisted thematic analysis
AI can be useful during thematic analysis when its role is defined narrowly. It can summarize a transcript for familiarization, suggest initial codes, identify passages that resemble an existing code, cluster reviewed excerpts into candidate groups, and propose possible theme labels. These are proposal functions. They should be treated as ways to accelerate inspection, not as automatic production of findings.
A rigorous AI-assisted workflow keeps each suggestion auditable. If AI proposes a theme called "autonomy with hidden developmental cost," the reviewer should be able to inspect the excerpts that support it, reject weak passages, edit the theme name, write a memo, and decide whether the theme belongs in the final analysis. The theme becomes part of the study only after human review. This is why OpenVerbatim uses a suggested-to-confirmed model: it aligns the interface with the methodological requirement that interpretation remains accountable.
The boundary is simple but important. AI can help you see candidates faster. It cannot tell you what your study means. It lacks your fieldwork context, your relationship to participants, your theoretical stance, and the ethical obligations attached to the data. Treat AI as a careful assistant whose work must be checked against source material, not as a substitute researcher.
Putting the workflow together
A practical workflow might look like this. Read the transcripts and write familiarization notes. Generate initial codes manually on a subset, then let software help you apply and retrieve them consistently. Use AI to suggest additional candidate codes only after you have enough grounding to evaluate them. Search for themes by comparing reviewed coded excerpts. Review each theme against the full dataset. Define the strongest themes in analytic language. Finally, write the report with quotes that can be traced back to confirmed evidence.
For hands-on practice, start with the interview transcript coding tutorial and download the sample transcript. Then compare the qualitative coding examples to see how open coding, axial coding, in vivo coding, and thematic coding differ. If you are evaluating tools, read the AI qualitative data analysis page, the open source qualitative research software guide, and the NVivo alternative comparison before trying the public sandbox.