How to code qualitative interviews with AI in the loop

How to code qualitative interviews: the choices that matter

How to code qualitative interviews is one of those questions whose right answer changes every few years. Before AI tooling matured, the answer was: a methodology textbook, a qualitative coding software license (NVivo, ATLAS.ti, MAXQDA), and weeks of manual work. After AI tooling matured, the answer is the same plus an AI layer that compresses the mechanical parts. The methodological choices have not changed; the time and cost of executing them have.

Coding is the systematic application of labels (codes) to spans of transcript text. The labels capture concepts, themes, or analytic categories. The same label may apply across hundreds of segments in a corpus, and tracking those applications is the central housekeeping task of qualitative analysis. Transcript coding software exists to make that housekeeping bearable; interview coding software is the same category seen from a different naming convention.

The choices that matter are: which methodology drives the coding (thematic, grounded theory, framework, content analysis), whether the codebook is generated inductively or applied deductively, whether the analysis is single-coder or multi-coder, and how AI assistance fits into the loop. Each choice constrains the others; the wrong combination produces a methodologically incoherent project.

AI assisted thematic coding: what changes

AI assisted thematic coding inserts a candidate-suggesting layer between the analyst and the corpus. Instead of starting with a blank codebook, the analyst starts with an AI-suggested set of candidate codes, organized by frequency and proximity in the data. The analyst then refines: merging duplicates, splitting over-broad codes, renaming for clarity, and adding codes the AI missed.

That refinement step is non-negotiable. Tools that present AI-suggested codes as final outputs produce homogenized analyses across teams — every project looks like every other project because the model regression-toward-the-mean dominates. Tools that present the suggestions as drafts and require human refinement preserve the analytical voice the methodology demands.

Automated thematic analysis and automated qualitative coding are the marketing-friendly names for these workflows. The honest framing is "AI-assisted" — the human is still doing the analysis, the AI is doing the secretarial work. AI thematic analysis software that hides this by default is making the wrong tradeoff for serious research.

Manual coding vs AI coding: the comparison researchers actually need

The right side of that table is not a replacement for the left side; it is a complement. Manual coding vs AI coding is rarely a binary — most projects benefit from manual on the first three interviews (to develop theoretical sensitivity), then AI-assisted on the rest (to scale).

Qualitative coding software vendors have caught on to this — most major tools now offer optional AI features rather than replacing the coding interface entirely. Thematic coding software with both modes lets the researcher pick per-project, which is the right product decision.

How to perform thematic analysis with AI in the loop

How to perform thematic analysis with AI assistance follows the same six-phase Braun & Clarke structure as manual analysis: familiarization, generating initial codes, searching for themes, reviewing themes, defining themes, producing the report. AI compresses phases 2 and 3 most; phases 1, 4, 5, and 6 are still human work.

Familiarization: read the transcripts. AI cannot do this for you, and the temptation to skip it produces visibly worse analyses. The qualitative interview summarizer pass helps the analyst familiarize faster — a 30-minute summary review is closer to "skimmed" than "read," but combined with spot-checks against full transcripts it is workable.

Initial coding: AI does the heavy lifting. The model proposes codes for every utterance, the analyst reviews, refines, and approves. Searching for themes: AI suggests theme clusters from the codes; the analyst rearranges, merges, and renames. Reviewing themes against the data is human work — the AI does not have a feel for whether a theme actually captures the data, only whether it is statistically prominent.

Defining and naming themes is human writing. So is the report. AI can draft the codebook reference and pull quote excerpts that support each theme, but the framing — what makes a theme matter, what tension it sits at, what it adds to the literature — is the analyst's contribution and cannot be outsourced.

Interview transcript coding examples and templates

Interview transcript coding examples are the fastest way to build intuition. A typical thematic analysis example: an interview about a participant career change might generate initial codes like "financial pressure," "identity shift," "family expectations," "skill mismatch," and "fear of failure." Those become candidate themes when they recur across participants — "the financial-identity collision" might emerge as a higher-order theme combining the first two.

Qualitative data analysis workflow templates are widely available — Saldaña has a coding manual, Braun & Clarke have published step-by-step guides, and most QDA software ships with sample projects. The AI layer does not replace these; it accelerates execution within them. A first project should use a published workflow rather than improvising.

For projects with very large corpora — how to analyze 50 interviews quickly is a common search query — the AI-assisted workflow becomes essential rather than convenient. Manually coding 50 interviews is a multi-month project. AI assistance brings it to weeks. The risk is that the analyst loses contact with the data; the mitigation is structured spot-checks at every phase.

AI grounded theory analysis: a special case

AI grounded theory analysis is the most contested intersection of AI and qualitative methods. Grounded theory commits to deriving codes from the data without imposing pre-existing categories. AI suggesting codes upfront violates that commitment in spirit, even when the analyst overrides the suggestions.

The legitimate use of AI in grounded theory is downstream of open coding: constant-comparison support (find similar quotes to this one), code-application mechanics after the analyst has done the open-coding pass manually, and memo drafting where the analyst writes and the AI summarizes. Code suggestion at the open-coding stage should be skipped if the methodological commitment is grounded theory.

That distinction matters for committee review. A grounded-theory dissertation that documents AI use only in the supporting infrastructure — not in the open-coding phase — survives methodologically. A dissertation that started with an AI-suggested codebook will not.

How to summarize research interviews accurately

How to summarize research interviews well is a skill in itself. The summary has to preserve what the participant said, what they emphasized, and what tensions or contradictions appeared, while compressing length by an order of magnitude. AI can draft the summary; the human review is what catches the inevitable hallucinations and emphasis errors.

A reasonable summary template: opening context (who, when, format, length), main topics (3-5 bullets with one supporting quote each), notable tensions (where the participant contradicted themselves or expressed ambivalence), unexpected mentions (what came up that the protocol did not anticipate), and analyst notes (where to follow up in subsequent interviews). AI handles the first three reliably; the last two need human attention.

For projects where the summary itself is a deliverable — internal stakeholder readouts, for instance — invest in human review of every summary. For projects where the summary is just analyst infrastructure (helping you decide where to focus), an AI-only summary with spot-checks is acceptable.

One useful pattern: ask the AI to summarize once before any human review, then summarize again after the analyst has read the transcript. The two summaries are usually different, and the differences highlight what the analyst noticed that the AI missed. That delta is itself an analytical artifact worth keeping — it is the closest mechanical proxy for what makes human qualitative analysis valuable.

For longitudinal projects with the same participants across waves, summary continuity is its own challenge. The summary of session 1 should inform the read of session 8 — the analyst remembers what mattered the first time, the AI usually does not. Tools that let you attach prior summaries as context to the next session AI pass produce noticeably more coherent longitudinal analyses than tools that treat each session as independent.

Summary length is the other under-discussed variable. A two-paragraph summary forces the AI to compress harder and surfaces the most prominent themes; a two-page summary is closer to a paraphrase and rarely earns its read time. Ask the AI for both lengths on the first few transcripts to calibrate; once the analyst has a sense of which length surfaces what they need, standardize on that length for the rest of the project.