Skip to content

How does AI thematic analysis work?

By Khyati Ojha · · 6 min read

How AI thematic analysis turns transcripts into themes, what research says it does well, where people still decide, and what to ask in a tool demo.

How does AI thematic analysis work?

You work on the rider support team at a bike-sharing company. This week, a vendor showed your team a demo. It turned a batch of rider complaint calls into a neat list of themes in under a minute. Now your manager wants to know how it got there, and whether to trust it.

AI thematic analysis follows the same path a person would, only faster. The tool splits each transcript into short excerpts. It suggests a code, a short label, for each one. Then it groups the codes into themes and names them. AI can come close to people on the big themes, but it slips more on single quotes. So trust a theme only when you can click from it to the quotes behind it.

Five steps explain where those themes came from, and five questions tell you what to check in a demo.

The short version

  • An AI tool cuts transcripts into excerpts, labels them, then groups the labels into themes.
  • A code labels one excerpt. A theme is a pattern across many codes.
  • AI is fast and steady on the first pass, and close to people on the big picture.
  • It's less steady on single quotes, so a person still checks the themes.
  • In a demo, click from each theme to its quotes, and from each quote to its transcript.
  • Bring your own transcripts to the demo, not the vendor's sample.

How AI turns transcripts into themes

Five steps an AI tool takes: the call becomes a transcript, the transcript is cut into excerpts, each excerpt gets a code such as "full dock", codes are grouped into a theme, and a person reviews the themes against the quotes.

AI thematic analysis turns interviews into themes in five steps. It makes transcripts, cuts them into excerpts, labels each excerpt, groups the labels into themes, and hands the result to a person to review. Each step can go wrong, so it helps to know which one made the themes on the demo screen.

1. Interviews become transcripts

The tool turns each recorded call into text. If this step drops words or mixes up who's talking, every later step builds on the mistake.

2. Transcripts are cut into excerpts

An excerpt is a short piece of text, usually one sentence or one answer. For the bike scheme, one excerpt might be "the dock was full, so I rode around for ten minutes."

3. Each excerpt gets a code

A language model reads each excerpt and suggests a short label, called a code. That excerpt might get the code "full dock." Some tools skip labels and group similar excerpts by meaning instead.

4. Codes are grouped into themes

The tool groups related codes and names each group. A code labels one excerpt. A theme is a pattern across many codes, like "riders can't end trips where they want to." Good tools show sample quotes under each theme.

5. A person reviews the result

The researcher merges, splits and renames themes. They also drop any theme the quotes don't support. Good tools keep every quote linked to its place in the transcript, so this check is quick.

What does AI do well in thematic analysis?

AI does the first pass of thematic analysis well. It codes and groups hundreds of excerpts in minutes, and it applies the same labels the same way every time. Research also finds AI can come close to people when you compare whole themes. That makes it useful for big piles of transcripts.

Speed is the clearest gain. In a 2025 study, 18 researchers tested an AI toolkit for thematic analysis. It sped up coding and theme building a lot, and the researchers still did the deeper reading themselves.

Agreement with people is improving too. A 2024 study compared AI coding with human researchers. With an updated model and a better prompt, the AI's results lined up more closely with the human coders, when you added up scores theme by theme.

There's also steadiness. A person coding call 200 is more tired than they were at call one. A tool applies the same codes the same way across the whole set.

Can AI do thematic analysis on its own?

No. AI can do the first pass, but a person still needs to decide what the themes mean. The same 2024 study found that while the theme totals lined up, the AI's labels on single excerpts varied more. So a theme can look right while some of its quotes are filed in the wrong place.

People also bring things a tool doesn't have:

  • They know the business question the themes must answer.
  • They notice the odd remark that turns out to be the real finding.
  • They catch jokes, irony and cultural meaning that change what a quote says.
  • They decide what the team should do next.

So build in one step before anyone acts. A researcher reads each theme against the transcripts before a decision rests on it. Say the demo shows a theme called "app crashes." Open five of its quotes. If two are really about a full dock, the theme needs fixing. The post on how AI is changing qualitative analysis goes deeper on where AI helps.

Questions to ask in an AI analysis demo

A made-up example of what a theme should show in a demo: the theme "Riders can't end trips where they want to" with three rider quotes beneath it, each with a button to open it in the transcript.

Every demo you see will now show AI. A 2025 paper in the journal Quality & Quantity reports that generative AI is built into the three best-known programs for analyzing interview data: MAXQDA, ATLAS.ti and NVivo. So the real question is how well a tool shows its work. Ask these five questions, and bring your own transcripts.

1. Can I click from a theme to every quote behind it?

A good answer is yes, in one click. Then click from a quote to its place in the transcript. If you can't, you can't check the theme.

2. Can I edit codes and themes, and do my changes stay?

You should be able to merge, split and rename. The tool should keep your edits when you add more interviews.

3. Does it work across languages and markets in one study?

If your riders speak several languages, ask to see themes built from more than one. Check the quotes in each language.

4. What happens to our data?

Ask where transcripts are stored and who can see them. Ask whether your data is used to train the vendor's AI models. Get the answer in writing.

5. How does it handle a theme it can't support?

Ask what happens when only one or two quotes fit a theme. A good tool shows how many excerpts sit behind each theme, so thin themes stand out.

If a tool passes all five with your own transcripts, it's worth testing on a real project. If it fails the first question, stop there.

Where Pulse fits

Pulse runs the interviews and the analysis in one place. Pulse runs AI-moderated voice interviews, and the report builds as each interview finishes, ready minutes after the last one. Interviews run in 240+ languages, enabled on request, and one study can cover several markets.

Pulse groups what people said into themes, each with the quotes behind it. Every quote links to the transcript it came from, so the first demo question has a one-click answer. Your researcher still reads the themes and decides what they mean.

To see it on your own study, book a demo, start on the $0 Explorer plan, or ask for a free study.

Questions readers ask

1. Which software is best for thematic analysis?

It depends on your work. Ask whether you also need to run the interviews, how many transcripts you have, and who will read the results. Then test each option with the five demo questions above, using your own transcripts. The right tool is the one that lets you check every theme quickly.

2. Is NVivo better than ChatGPT for thematic analysis?

They do different jobs. NVivo keeps codes, notes and sources together for a whole project. A chatbot can help with a quick first pass on a few transcripts. But it doesn't keep a project, and it won't link each quote back to its source for you.

3. How do you show a thematic analysis?

Use a simple table. List each theme with a one-line definition and two or three quotes. Link each quote to its transcript. Add how many people raised the theme, as a count, not a percentage. That gives readers the finding and the proof side by side.

4. How do you check that AI themes are backed by real quotes?

Pick a few themes and open every quote under them. Check that each quote says what the theme claims. Then open the transcript around each quote, to make sure the meaning holds in context. If many quotes don't fit, rebuild that theme by hand.

Book a demo

Book a Demo

Your details

By clicking “Book Now”, you agree Pulse may use your information as described in our Privacy Notice, including to contact you about our products and services.