Thematic analysis is how you turn a pile of transcripts into something a team can act on. You find the patterns of meaning that run across your qualitative data. Done by hand, it is brutally slow. You read every transcript, tag passages, group codes, refine themes. For decades that slowness capped how much qualitative data a team could actually digest.
AI lifted the cap. By 2026, the large majority of researchers use AI somewhere in their qualitative analysis, and AI-assisted coding cuts the time it takes by roughly two thirds while keeping the rigor intact, provided you use it correctly. That last clause is the whole game. Use it well and you get a serious accelerator. Use it carelessly and it will invent themes that are not in your data. This guide is the workflow for the first outcome.
The six steps of thematic analysis, accelerated
Thematic analysis still follows the same six steps it always has. AI speeds up several of them, but you still have to walk through all of them.
Step 1: Familiarize yourself with the data
Read the transcripts. Yes, even with AI helping. This is the step people are most tempted to skip, and the one you should not, because your own immersion in the data is what lets you catch it later when the model proposes a theme that is not really there. AI can give you a fast summary of each session to orient you, but a summary is a map, not the territory. Spend real time in the raw material.
Step 2: Generate initial codes
This is where AI first earns its keep. Coding, which means labeling the relevant passages, is the most laborious step, and AI can propose an initial set of codes across your whole dataset in minutes instead of days. Let it do the first pass: "highlight everything related to pricing confusion," "tag every mention of a workaround."
Then treat those codes as a draft, not a verdict. Read them against the transcripts and correct, merge, and cut. The model is fast and tireless and occasionally confidently wrong, so your judgment is the filter.
Step 3: Search for themes
Group related codes into candidate themes. AI can cluster codes and suggest where the patterns lie, which helps a lot on a large dataset where you cannot hold all the connections in your head. Here too you are the editor. The model proposes the grouping, and you decide whether each theme is real and meaningful or just a surface-level co-occurrence.
Step 4: Review the themes
Check each candidate theme back against the data. Does it actually hold across the transcripts, or did a couple of vivid quotes make it look bigger than it is? This is where your familiarity from step one pays off. If the model surfaced a theme and you cannot find solid support for it when you go back to the source, that is exactly the kind of plausible-but-wrong output to cut.
Step 5: Define and name the themes
Write a clear definition for each theme and give it a precise name. This part is human work. The name and definition encode what the theme means and why it matters, and that judgment is yours. AI can draft, but the final framing should be something you stand behind.
Step 6: Write up the findings
Produce the narrative, with themes supported by real quotes and examples. AI can assemble first drafts and pull supporting quotes quickly, but keep every quote anchored to a real participant, and lead the write-up with what the findings mean for the business, as we describe in translating insights into business metrics.
The one rule that prevents hallucinated insight
The central risk of AI analysis is the model producing themes that sound plausible but are not actually grounded in your data. There is a known tension here worth understanding.
You can constrain the model by defining categories in advance and having it code only into those buckets, which sharply reduces hallucination. But doing that defeats one of the main purposes of thematic analysis: surfacing emergent themes you did not anticipate, the surprises that make qualitative research valuable. Lock it down too hard and you only ever confirm what you already thought to look for.
The workable balance is to let AI propose codes and themes openly so you keep the emergent ones, then verify every theme against the source data before you trust it. Keep the freedom to discover and the discipline to check. Do not present a theme the model surfaced until you have personally traced it back to real quotes.
Why real data is the precondition
All of this assumes the data underneath is real. Run AI analysis on synthetic users and you get sophisticated processing of fabricated input, which is worse than useless because it looks rigorous. The workflow above is for analyzing real conversations with real participants, where the themes you verify are grounded in something true. Strong analysis starts with strong collection, which is why how you run the interviews matters as much as how you analyze them. For the manual fundamentals that still underpin all of this, see our detailed guide on how to synthesize qualitative data.
Where this fits at User Evaluation
User Evaluation runs collection and analysis in one place. AI-moderated interviews feed directly into AI-assisted synthesis, so the coding and theme-finding happen on real participant transcripts, with every theme traceable back to the exact moment a participant said it. That traceability is what lets you move fast on the coding while still verifying before you trust, which is the point of doing thematic analysis with AI rather than just letting AI do it for you.
What this actually buys you
AI makes thematic analysis far faster, pulling the slowest steps from days down to hours, but it does not change what the method is. Stay immersed in the data, let AI draft the codes and themes, and check every one against the source before you trust it. You keep the freedom to find emergent themes and the discipline to ground them in real quotes. Work that way and you get the speed of automation together with the rigor of real analysis, instead of giving up one to get the other.