How to Conduct Thematic Analysis with AI

What is Thematic Analysis (and Why Use AI?)

Thematic analysis is the process of analyzing qualitative data (such as interview transcripts, user feedback, or social media comments) to identify repeating themes. Usually, this process involves a series of steps where researchers familiarize themselves with the data, generate initial codes, iteratively review and refine those codes, and finally interpret the themes to generate insights. Thematic analysis is a popular approach in UX research and market research because it can reveal what people are really thinking and feeling. For example, a UX team might look at app review comments to identify common problems, or a marketing team might want to analyze open-ended survey responses to understand how customers feel about different products.

Manual thematic analysis can be very time consuming and subjective. This is where AI can help. AI thematic analysis software can assist with each of those steps using techniques such as natural language processing (NLP). It can quickly scan large volumes of qualitative data, automatically suggest codes or categories, group those codes into themes, analyze sentiment, and even generate reports – at a speed and scale that would overwhelm humans. In other words, AI acts as a programmable research assistant that can help you arrive at insights faster, while reducing some forms of human bias and error in the process.

Benefits of AI Assisted Thematic Analysis

Why should you use AI for thematic analysis? There are a number of benefits to using AI tools for thematic coding and analysis. Here are some of the main ones:

Efficiency: AI can process and code large datasets in a matter of minutes or hours, automatically identifying potential themes and helping to organize the data, which can massively reduce your analysis time. Whereas manual coding might take weeks to get through hundreds of survey responses, AI could deliver initial themes in a day or less.
Consistency: Humans get tired, make mistakes, and get off-topic – AI doesn’t (okay, at least the tired bit). AI automatically and consistently applies the coding rules you’ve defined across all your data, reducing variation in how codes are applied. That can lead to more reliable and repeatable results, especially on large projects.
Insight: Good AI thematic analysis software will also help you identify patterns you may not have thought of. It might detect less obvious but still frequent recurring phrases or links in the data that you might have overlooked. By analyzing every single word in your data, AI ensures nothing is missed, potentially surfacing insights you hadn’t considered.
Sentiment: Many AI tools also provide sentiment analysis in addition to thematic coding. As the AI system identifies themes, it also tells you the sentiment (positive, negative, neutral) people are expressing about those themes. For instance, if the AI finds that “customer support” is a theme, it might highlight whether responses that refer to customer support are mainly praising it or complaining about it.
Visualization and Reporting: AI doesn’t just identify themes – it can help you present them, too. Some platforms auto-generate thematic visualizations, like word clouds, or reports with charts, tables, and descriptive text summarizing the findings. This makes it easier to present insights to stakeholders. For example, AI might output a list of key themes with color-coded tags, the percentage of responses that fall into each theme, and example quotes, ready to copy into your report.

Essentially, AI can take care of the hard work of thematic analysis while you focus on the fun part of interpreting results and deciding what to do next. But how do you actually conduct an AI-assisted thematic analysis? Let’s take a look at it in more detail, step by step.

Step-by-Step Guide: AI Thematic Analysis

Conducting a thematic analysis with AI follows similar steps to the manual process – you’ll code and develop themes, but with AI as a helpful sidekick. Below is a step-by-step guide, blending best practices in qualitative analysis with AI tools and techniques.

Step 1: Prepare and Import Your Data

Start by gathering your qualitative data – interview recordings, survey responses, user testing notes, or customer reviews. If your data is audio or video (like user interviews or focus groups), transcribe it so it’s in text form for AI tools to process. Modern tools can transcribe recordings quickly. For example, AI transcription features convert interview recordings to text with timestamps linking each phrase to the audio. Always double-check transcripts for errors, especially with accents or technical terms, but this beats manual transcription.

Next, clean and organize the data. Remove or anonymize personally identifiable information if needed (some AI tools offer automatic de-identification features). Combine your text data into the format your AI tool expects – uploading files to qualitative analysis software or pasting text into an AI assistant. At this stage, consider your research questions or objectives, as giving the AI context helps it focus on relevant themes.

Pro Tip: Some research platforms let you input research questions or project goals before analysis. Giving the AI this context helps it focus on relevant themes rather than tangents. For instance, if you’re discovering customer pain points around a mobile app, tell the AI upfront so it can look for themes around problems and frustrations.

Step 2: Familiarize Yourself (and the AI) with the Data

In manual thematic analysis, researchers first familiarize themselves with the data by reading transcripts and getting a general sense of what’s been said. This is still valuable, but AI can help summarize and explore the data upfront.

Most AI-powered thematic analysis tools have features that automatically summarize transcripts or highlight key points. Upon uploading data, the tool might generate a short summary of each interview or show frequently mentioned topics. Some UX research tools have AI summary features that identify key moments from interviews – like quotes or insights – displaying them alongside the transcript. You can scan these AI-generated summaries to orient yourself, but take care not to overtrust them. They’re useful guides, but spot-check the actual data to ensure nothing important was misinterpreted or omitted.

You can also use AI for keyword extraction or rough clustering at this stage. Some tools scan text and show a word cloud or frequency list of terms. The software might lump together similar responses based on word similarity – a warm-up, automatic version of initial coding. For instance, AI might reveal that words like “price”, “quality”, and “customer service” appear frequently in feedback data, flagging potential theme topics to explore later.

Step 3: AI-Powered Coding of the Data

Now we reach the core of thematic analysis: coding. Coding involves labeling pieces of text – phrases, sentences, paragraphs – with tags or “codes” that describe their content. Traditionally done by reading line by line and applying codes manually, AI can help speed up thematic analysis coding.

AI can help with coding in a couple of ways:

Auto-Coding inside Software: Qualitative analysis programs like BTInsights can automatically code your text data in minutes instead of days or even weeks. For example, the BTInsights survey open-ends coding tool can code survey text with human-level accuracy. You input your text, and the software outputs a set of codes it applied to parts of the data – essentially like having a junior researcher do an initial pass of coding.
Use a General AI (e.g. ChatGPT): Without access to specialized software packages, you can use something like ChatGPT to help code data. You may need to break data up due to input size limits, but it’s as easy as copying and pasting response text (like answers to open-ended survey questions) and asking, “What are the key themes or codes in this text?” ChatGPT can then generate a list of possible codes or summarize the content. AI outputs possible relevant codes within seconds. For example, given a customer comment like “The app is easy to use but kept crashing during checkout,” it might output codes like “ease of use” and “app crashes/bugs.”

Go through AI-suggested codes and edit them. Merge similar ones, delete those that don’t make sense, and add important codes the AI missed. This is where your domain expertise matters. For example, AI might give you separate codes “price” and “cost” which you realize are the same thing – so you consolidate them into one “Pricing concerns” code. You may notice themes in the data the AI completely missed, so you can manually code those pieces of text. (Perhaps few people mentioned “customer support” and because AI looks for common patterns, it overlooked it, but you know it’s important for the analysis.)

Think of AI as having done the grunt work of initial coding. It gives you a rough draft of codes to work with. You as the analyst are like an editor reviewing that draft, finding AI’s mistakes or weak decisions. This human-in-the-loop approach is important because AI, while fast, doesn’t truly understand nuance or context the way humans can.

Step 4: Identify Themes Using AI Clustering

Coding is well underway at this point. Next, start grouping codes into larger themes. A theme is a higher-level pattern in the data that captures something important. If codes are individual tags, a theme is like a category that unites several codes. For example, you might have codes like “slow load time”, “app crashes”, and “buggy notifications” that all fit under a higher theme of “Technical Issues”.

AI can help with clustering your codes or excerpts into themes. Most tools have some way of automatically grouping similar codes. For example, AI might notice several codes about price, cost, and discounts that co-occur often, and automatically suggest they could be clustered into a tentative theme like “Pricing and Value.” UX research software with AI can present groups of quotes discussing the same underlying concept. In this example, AI has automatically clustered dozens of comments to provide a quick way to see potential theme groupings (though it may dump unrelated bits in an “Other/Outliers” cluster, which is common).

Without these UX platforms, you can do something similar using a general AI like ChatGPT. Provide it with your set of refined codes and ask it to organize or categorize them. For example: “Here are some codes from my data: {list of codes}. Can you group these into 3-5 broader themes and suggest a name for each theme?” ChatGPT might respond with something like: Theme 1: User Interface (codes: navigation, layout, design); Theme 2: Performance (codes: slow load, crashes, lag); etc. This can jumpstart ideas for structuring your themes.

Step 5: Review and Refine Themes (Human in the Loop)

With a list of themes in hand, you can now review and refine them. This is where your human involvement is at its maximum. AI can accelerate the coding and even suggest groupings/clusters, but only you can validate that the themes make sense, are unique, and accurately represent the data.

Review the original quotes, notes, etc., and read the content of each theme. Does that content really fit the theme? Did anything get missed? Is anything coded here that shouldn’t be? If you find something fishy (like a comment about “pricing” that the AI stuck in the “usability” theme), move it to a different theme or recode it. Also check the opposite: did you miss anything? Did the AI collapse two related ideas into one theme and you realize they should be separate? You can split a theme into two, or two themes into one if they turned out to be different facets of the same problem.

Remember that thematic analysis is an iterative process: it’s normal to jump between coding and themes a few times. AI tools will typically have some process for reprocessing (or iterative passes): for example, once you have done some manual fixes of codes or theme groupings, you can run an AI “recheck” to see if any new patterns emerged. Some AI qualitative analysis tools will do multiple iterations for you, merging duplicate themes and refining definitions each time. Use that feature if it’s available, but sanity-check the results, of course.

While you are reviewing, consider having a colleague or another team member review too (if possible). Human peer review will help check for biases or blind spots that both you and the AI have missed. For example, maybe you (and thus the AI, if it was trained or biased by you) were looking for negative feedback themes so hard that you missed an area of strong positive feedback. A second set of human eyes can spot that.

Once this review/refinement step is done, you should have a strong set of themes that you are happy with and confident in. Each theme should be clearly defined, and you should have some example quotes/data excerpts for each that show what it means (you can have the AI pick out example quotes if you like: e.g. “AI, give me one or two example quotes that best illustrate the X theme.” It’s usually good at finding them).

Step 6: Interpret and Report the Themes

The final step is to make sense of the themes and share the insights.

Interpret the themes in the context of your project and audience. That means understanding and communicating the data’s implications (themes + sentiment) in the context of the business/marketing/UX design/product context in your head. A marketer might see a theme around “Brand Trust” in a customer feedback analysis connecting to recent PR issues and the need for a campaign to address it. A UX designer might see themes in user research, like “Navigation Confusion”, pointing to specific product design issues to solve. AI can’t understand your company’s specific context or strategy – that’s for you to interpret.

As you put together your final findings, consider telling a story/narrative: Why do these themes matter? What should we do about them? That’s where you demonstrate the analysis’s value. AI did most of the legwork, but you deliver the insights.

AI Tools and Software for Thematic Analysis

By now you might be wondering, “Which tools can I use to do all this AI magic?” There are many thematic analysis software options incorporating AI today. Here we’ll mention a few categories and examples (and don’t worry, we’re not sponsored by anyone – just sharing for your exploration):

AI-Native Thematic Analysis Platforms

BTInsights is one of the best AI-native platforms for thematic analysis. It can extract thematic themes from interview or conversation data. It can also automatically generate codes for open-ended survey responses. AI-native platforms for qualitative analysis like BTInsights not only have easy to use interfaces but also return far more accurate results than traditional qualitative analysis software. The difference is that the former are built from the ground up with AI as their foundation while traditional platforms attempt to shoehorn AI into already existing structures.

Key features of the BTInsights thematic analysis platform:

Interview transcription and translation
Thematic analysis
Find quotes
Create video clips and highlight reels
Code open-ended survey responses
Create cross-tab tables
Generate PowerPoint slides in your own format

Traditional Qualitative Analysis Software with AI Features

Programs like NVivo and ATLAS.ti have long been used for manual qualitative coding and now offer AI enhancements. NVivo, for instance, can auto-code text to suggest themes and speed up your analysis. ATLAS.ti similarly bridges human expertise with AI efficiency, allowing automatic coding based on your intent and even AI transcription built in. These are robust tools ideal if you regularly do qualitative research, though they can be a bit pricey and have a learning curve.

General AI Assistants

You can always use general AI models like OpenAI’s ChatGPT (or GPT-4) as a flexible tool for thematic analysis. While ChatGPT isn’t a dedicated thematic analysis software, it can handle many tasks if you prompt it well – from generating initial codes to categorizing comments. It’s essentially a do-it-yourself approach: you paste in data (in chunks) and ask for analysis. The advantage is it’s very easy to use (no special software needed). The downside: it can struggle with large datasets (due to input limits), and it may occasionally produce incorrect or nonsensical outputs if not guided carefully. For small projects or quick brainstorming, though, it’s a handy option.

When choosing an AI thematic analysis tool, consider factors like cost, data security, customization, and integration with your workflow. A UX team might value a tool that integrates with their interview recording software and has video clip capabilities, whereas a market researcher analyzing survey comments might prefer a tool that can handle thousands of responses with robust text analytics. You don’t have to stick to one tool either; sometimes a combination works (e.g., using ChatGPT to refine some definitions even if you got the themes from another software).

Best Practices When Using AI for Thematic Analysis

Before we conclude, let’s quickly run through a few best practices and cautions when doing thematic analysis with AI:

🤖 Be Transparent and Ethical: When using participants’ data with AI tools (notably cloud-based tools), make sure you have the necessary consent and are complying with privacy laws. Be explicit in consent forms about using AI for analysis if necessary. Check if the tool discloses or resells data to third parties, or uses the data to train their models, and de-select/opt-out/choose a privacy-friendly option if necessary (some enterprise tools allow you to process data locally, or choose which server/region to use).
⚖️ Be on the lookout for Bias: AI models inevitably bring along the biases of the data they were trained on. This might lead them to systematically underrepresent certain viewpoints or demographics in your analysis (say, if your data set includes slang or dialect not present in the AI’s training data, it will misinterpret or ignore those responses). You can’t know exactly where the bias might be, so combat it by double-checking for signs of systematic skew in your themes, and by keeping an eye out for and taking action against bias in any machine learning models you train yourself. You can also reduce unintended bias in an AI system by using systems that allow you to supply domain-specific context (allowing the system to tailor to your content better).
🕵️♀️ Don’t Abandon Ship – Be Involved: It can be all too easy with AI to import or upload data into an AI tool, run the algorithm, and take the result at face value. Avoid this temptation. As Nielsen Norman Group researchers succinctly advise, treat AI’s coding as an initial pass and remember that a human still needs to make sense of the data and connect the dots in ways AI can’t. An AI may spit out themes for you, but only a human can make sense of why those themes are the key themes in context. The AI also won’t have a clue about sarcasm, humour, or other subtleties of human communication – whereas you will, especially if you have familiarity with the participants or the domain. Audit the AI’s work.
🔄 Iterate: Use AI in iterative ways where possible. For instance, run auto-coding, then refine the codes, then maybe run the AI cluster again on the refined codes, and so on. Each iteration may improve the output. If the AI allows you to give feedback (some tools learn from your edits within a project), use that to get improved suggestions. It’s somewhat like training an assistant – the first set of suggestions might be terrible, but with more clear instructions, the assistant can start to meet your expectations.
📊 Use Quantitative as well as Qualitative Data where possible: This is more a general best practice for research in general, but given many AI tools make it much easier to combine qualitative and quantitative data. If you have quantitative metrics (ratings, usage analytics, etc.) in addition to your themes, compare and contrast them. For instance, AI may help by automatically clustering participants or responses not just by theme but by sentiment, usage levels, etc. A more complete picture can yield more rich insights (e.g., “Users who gave us low satisfaction ratings talked mainly about Theme X and Y – suggesting these themes are behind dissatisfaction with the product”).

By following these best practices, you’ll make sure your AI-assisted thematic analysis is not just fast, but also credible, accurate, and insightful. The goal is to leverage AI’s strengths (speed, scale, pattern recognition) while compensating for its weaknesses (lack of true understanding, possible but hidden biases) through human expertise.

Conclusion

AI thematic analysis is a great way to help make sense of unstructured qualitative data – and to do so much faster, and in some ways, more intelligently. In the context of research, the applications of AI to thematic analysis allow market researchers, UX designers, and marketers to identify and act on patterns and themes in user feedback or interview transcripts with speed and consistency. By automating the most time-consuming parts of thematic analysis coding and theme detection, AI allows you to focus on interpreting the data and acting on your findings.

So, the next time you have a set of user interviews to analyze for your UX research, or a batch of open-ended survey responses to comb for your product launch, consider bringing AI into the mix. Give it a try on a small sample to see what it can do (feel free to ask ChatGPT to suggest themes for a few user responses) or experiment with a free trial of a qualitative data analysis software that has AI capabilities built-in. You might be surprised at how quickly you get a useful first draft of an analysis. As always, the same rules apply as for any research process – question and verify, and put the findings into context.

BTInsights, as one of the best AI thematic analysis tools, can automate your entire interview and survey analysis from raw data to report, turning weeks of analysis work into hours or even minutes. AI-powered thematic analysis capabilities will give you a significant competitive advantage in a world where data (especially unstructured data like text feedback) is growing exponentially.