Using AI for thematic analysis in qualitative research

Key takeaways

• Create your codebook first: Define each code and include clear examples before giving your data to AI. This helps prevent the tool from grouping unrelated responses under the same theme.
• Test AI coding on a sample: Have a human review 10% to 20% of the AI-coded responses before analyzing the entire dataset. This gives you a chance to catch inconsistencies early.
• Use AI to manage large datasets: Typeform forms received 528.1 million submissions over the past year, including 178.4 million submissions from forms with a long-text question. At that scale, manually coding every response isn’t practical.
• Account for survey drop-off: Completion rates can vary considerably depending on the type of form. Employee feedback forms had a 74.53% completion rate, compared with 51.65% for employee recruitment forms. Invite enough participants to reach saturation even if some don’t complete the survey.
• Let AI code, but keep humans in charge of interpretation: AI can handle the initial coding process, but understanding what the themes mean and why they matter still requires human judgment.

Thematic analysis, or finding patterns and meanings in qualitative data, is one of the most powerful tools researchers have. It turns open-ended responses, interview transcripts, and unstructured feedback into actionable insights. But it’s also one of the most time-consuming parts of research.

Manually coding hundreds or thousands of responses to identify recurring themes requires patience, consistency, and hours of focused work. Researchers often find themselves rereading the same passages, second-guessing their category assignments, and struggling to stay objective as fatigue sets in.

AI is changing things. Modern AI tools can help researchers code faster, spot patterns humans might miss, and reduce the cognitive load of analysis, freeing up time for what AI can’t do: interpreting findings and connecting them to real business outcomes. By automating the initial coding phase, researchers can focus their expertise where it matters most, which is validating results and drawing meaningful conclusions.

Here’s what you need to know about using AI for thematic analysis, why it works, and how to make sure your results stay rigorous.

What thematic analysis does

Thematic analysis is the process of identifying, coding, and organizing patterns in qualitative data. Unlike quantitative methods that measure frequency or correlation, thematic analysis asks: what themes emerge when we look across many responses?

A product team might run interviews with 15 users about their onboarding experience. The researcher reads every transcript, marks sections where users mention “confusing” or “unclear,” then groups those mentions into a theme, say, “lack of clarity in setup.” Another theme might be “slow load times.” A third, “helpful customer support.” By the end, the team has a clear picture of what’s working, what needs improvement, and how users describe their experience in their own words.

This kind of analysis works across any qualitative source: interview transcripts, open-ended survey responses, customer feedback, social media comments, support tickets, or focus group notes. The output is always the same: a structured understanding of what your data contains. The richness of the method lies in how it preserves respondent language while organizing it into meaningful categories that reveal underlying patterns.

But here’s the challenge: thematic analysis requires reading through data multiple times, applying consistent definitions to codes, and staying alert to nuance. One person might code “the setup took a while” as “slow load times.” Another might code it as “confusing process.” Both are doing thematic analysis, but inconsistently, and that undermines the reliability of findings.

This inconsistency, sometimes called “coder drift,” becomes more pronounced as the analysis progresses and fatigue accumulates.

Scale that to 50 interviews or 500 survey responses, and the work becomes enormous, and so does the risk of inconsistency. The sheer volume of data makes it nearly impossible for a single person to maintain consistent coding standards throughout an entire project.

Why AI is suited for thematic analysis

AI doesn’t replace the judgment calls that make thematic analysis meaningful. It removes the grunt work that makes it slow.

Here’s where AI helps:

‍Speed in initial coding. AI can process large volumes of text in minutes instead of days. Once given coding rules or examples, it can apply those rules consistently across all your data. Researchers using AI for thematic analysis report that the tool handles initial coding in a fraction of the time manual coding would take. This acceleration opens possibilities for more ambitious research projects that would otherwise be constrained by time and resources.

‍Consistency across codes. AI doesn’t get tired, distracted, or biased by what it coded yesterday. It applies the same logic to the first response and the thousandth. That doesn’t make AI perfect. While it can still miss nuance or misinterpret context, it does reduce drift, which is a real problem in large qualitative projects. This consistency becomes increasingly valuable as datasets grow larger, since human consistency typically degrades over time.

‍Pattern spotting. AI can flag connections between codes that a human reviewer might overlook, especially in large datasets. If “confusing,” “steep learning curve,” and “unclear instructions” are coded separately, AI might suggest they belong under a broader theme. Again, the researcher decides, but the tool surfaces possibilities that might otherwise remain hidden in a large corpus of data.

‍Faster iteration. Because AI does the initial pass quickly, researchers can refine their code definitions and re-run the analysis multiple times. That iteration leads to better, more rigorous findings. The ability to test different coding schemes without enormous time investment fundamentally changes how researchers approach their work.

‍Transparency and auditability. When AI codes your data, you have a record of how and why each code was applied. That’s valuable for peer review, compliance, or simply checking your own work. This documentation creates accountability that manual coding often lacks.

More than half of researchers (53%) now regularly use AI in their work (Qualtrics, 2026 Market Research Trends).

In other words, the shift isn’t coming. It’s already here.

How to use AI for thematic analysis without losing rigor

AI is powerful, but thematic analysis still requires human judgment. Here’s how to keep your analysis rigorous.

Start with questions easy to code

AI can only identify patterns based on the responses your questions generate. If your questions are too broad, you’re likely to get vague answers that are difficult for both people and AI to categorize.

A few simple practices can help you collect cleaner, more useful open-ended responses from the start:

• Focus on one idea at a time. Split “What did you like and dislike?” into two questions so respondents don’t have to address two topics at once.
• Make questions specific. “What almost stopped you from signing up?” is more likely to produce useful insights than a broad question like “Any feedback?”
• Combine closed and open-ended questions. Ask respondents for a rating, then follow up with, “What’s the main reason for your score?” The rating adds context and makes it easier to compare themes across different scores.
• Use logic for relevant follow-ups. Show follow-up questions only when they apply to the respondent’s previous answer. This keeps the survey relevant and avoids asking people to explain answers they never gave.
• Collect background details automatically. Use URL parameters to capture information like plan, segment, or source without adding more questions. This makes it easier to compare themes across groups while keeping the form concise.

Define your codes before you code

Don’t just feed AI your raw data and ask it to “find themes.” That’s too loose. Instead, start with a codebook: a set of definitions for the themes you’re looking for, with examples of what qualifies.

For instance:

• Navigation confusion: Statements where users mention difficulty finding features, unclear menu structures, or taking longer than expected to locate something.
• Feature gap: Comments where users say they expected a feature but it doesn’t exist, or an existing feature doesn’t work the way they expected.
• Performance issues: Reports of slow load times, freezing, crashes, or other technical problems.

AI has guardrails when you’re this specific. It knows what you’re looking for, so it won’t conflate “slow” (performance) with “hard to find” (navigation). Detailed code definitions also serve as a communication tool, so stakeholders understand exactly what each theme represents and why specific responses were assigned to particular codes.

Give AI clear instructions for coding responses

Here is a sample prompt you can customize:

You are analyzing open-ended survey responses and assigning codes based on the codebook below:

[Code name] - [Definition] - [1-2 example quotes]

For each response:

1. Assign one or more codes from the codebook. If none apply, label the response “Uncoded.”
2. Include the exact words from the response that support each assigned code.
3. Flag any responses that are sarcastic, unclear, or unrelated to the question.

Use only codes from the codebook. Don’t create new codes or paraphrase respondents’ words.

Responses: [ID] - [Response text]

Requiring AI to include the exact words that support each code makes it easier to review its choices and quickly confirm whether each code was applied correctly.

Validate AI codes against a sample

Before you run AI coding across all your data, pull a small sample (10% to 20%) and have a human coder review it. This catches problems early. If the AI is misinterpreting your code definitions or missing obvious mentions, you’ll spot it before processing thousands of responses. This validation step acts as a quality gate to ensure that systematic errors don’t propagate across your entire dataset.

Stay involved in theme refinement

After AI does the initial coding, review the results. Look for patterns it flagged. Spot-check coded responses to make sure they match the theme. Merge codes that should be combined. Split codes that are too broad. This is where your domain expertise shines: you understand context, industry norms, and what your stakeholders care about in ways AI doesn’t. Your role transforms from data entry to strategic interpretation, where you can add genuine value.

Track sample size and saturation

Thematic analysis is done when you’ve reached saturation, which is the point at which new data stops revealing new themes. For research to be credible, you need to know when you’ve hit that threshold.

Near saturation (capturing 90% of codes) typically occurs at 15 to 23 interviews. Studies that use more structured interview guides and predefined codes often reach saturation more quickly (Journal of Medical Internet Research, 2024).

If you’re collecting qualitative data through open-ended surveys instead of interviews, focus on completed responses rather than the number of invitations you send. Our data shows that completion rates stayed between 59% and 61% each month over the past year, although rates can vary significantly depending on the context. Employee feedback forms, for example, had a 74.53% completion rate, compared with 51.65% for employee recruitment forms. Rates also differed by industry, with healthcare forms at 66.69%, compared with 54.24% for marketing and advertising. If your goal is 30 completed responses and you anticipate a 60% completion rate, plan to invite at least 50 participants.

AI doesn’t change these numbers, but it does make it easier to reach them. You can code more interviews in less time, which means you can say with confidence that you’ve interviewed enough people. This accelerated pathway to saturation removes a major constraint from qualitative research planning.

Document your methodology

Write down the code definitions you used, how you validated AI coding, which tool you used, and any adjustments you made to the AI’s output. This transparency, which is sometimes called “explainability,” is essential for credibility. Reviewers, stakeholders, and future researchers need to understand how you arrived at your conclusions. Documentation also serves a practical purpose: it makes your work reproducible and helps you remember your reasoning months later.

Tools and workflows for AI-assisted thematic analysis

Many research platforms now include AI-powered coding and theme identification. When evaluating a tool, look for:

‍Customizable code definitions. Can you input your own codes, or does the tool force you to use pre-built themes? The more flexible the tool, the better it fits your specific research question. Tools that allow customization give you control over what constitutes a valid interpretation of your data.

Explainability. Does the tool show you why it assigned a code? Can you see the exact text it’s responding to? A black box is less useful than a tool that lets you audit its decisions. Explainability builds trust and helps you catch errors before they compound across your dataset.

‍Integration with your data collection. If you’re running surveys or interviews within the same platform, you want coding and analysis built in. Switching between tools slows down your workflow and risks losing context. Integrated workflows keep your research contained in one place, reducing friction and maintaining continuity.

Speed and accuracy trade-offs. Some AI tools prioritize speed, while others prioritize precision. Understand which you need. If you’re doing exploratory research where speed matters, a faster tool might be better. If you’re doing compliance or clinical research where accuracy is non-negotiable, a slower, more careful tool is the right choice.

Tools like ResearchFlow are designed to keep qualitative research in one place, from collecting responses to generating themes, summaries, and sentiment, without switching tools. That continuity matters more than it might sound. It reduces friction when you iterate on your findings. When you use AI thematic analysis in an integrated platform, you get insights as responses come in and can quickly pivot your coding strategy if early results suggest a different approach.

Common pitfalls to avoid

Trusting AI without verification. AI is fast, but it’s not always right. It can miss sarcasm, misread context, or apply codes too broadly. Always spot-check. Always have a human review high-stakes findings. Verification isn’t a sign of distrust. It’s a best practice that ensures quality.

Using quotes without verifying them. AI may paraphrase a respondent’s words or combine parts of different responses into a quote that was never actually said. Before including any quote in a presentation or report, check it against the raw data to make sure it’s accurate.

‍Skipping code definition. The more precisely you define your codes, the better AI performs. Vague definitions lead to vague results. Spend time upfront on this. The investment in clarity at the beginning saves time and improves accuracy downstream.

‍Ignoring the sample size. AI speed is tempting. You might code 8 interviews and call it done. But saturation still matters. Smaller samples miss themes. Use AI to make larger samples feasible, not to justify smaller ones. The goal is rigorous analysis, not fast analysis.

‍Losing the qualitative richness. Thematic analysis is powerful because it preserves quotes, context, and nuance. Don’t let AI efficiency tempt you to strip that away. Your findings should still be grounded in actual respondent language. Rich, quoted findings resonate with stakeholders far more than abstract theme lists.

Adding personal data to AI tools. Open-ended responses can contain sensitive information like names, email addresses, or account details. Remove any identifying information before using AI to analyze the responses. It’s also important to know how your AI tool handles uploaded data, including whether it stores your information or uses it for training. This matters most when you’re working with customer or employee feedback.

‍Forgetting that coding is interpretation. There’s no perfectly objective code. Every definition you write is a choice about what matters. AI doesn’t change that. It just makes your choices visible and scalable. Own that responsibility.

Why this matters now

Qualitative research is becoming more common, not less. Surveys remain the most widely used quantitative method among market research professionals, with 85% using them regularly (Backlinko, 2026). But organizations increasingly recognize that understanding the reasons behind the numbers requires deeper insights, and that’s where qualitative data comes in.

At the same time, organizations are collecting qualitative data from more sources than before, including open-ended survey responses, customer interviews, support feedback, and social listening data. On Typeform alone, forms received 528.1 million submissions over the past year. Forms with a long-text question, the open-ended format thematic analysis often starts with, received 178.4 million submissions on their own. Analyzing that amount of data manually isn’t practical.

AI makes that volume more manageable. It makes it possible to do rigorous, large-scale thematic analysis in weeks instead of months. That means you can get insights faster, iterate faster, and make decisions with confidence.

But speed is only valuable if it doesn’t sacrifice rigor. That’s why the human element remains central. AI codes, you interpret. AI flags patterns, you validate and refine. AI scales analysis, you stay responsible for what it means. Using AI thematic analysis doesn’t diminish the researcher’s role. It elevates it, removing tedium so expertise can shine.

The future of thematic analysis isn’t AI replacing researchers. It’s researchers using AI to focus on what they do best: making sense of human insight.

Source: Typeform internal data from 882,461,930 form starts across approximately 2.4 million forms, September 2025 to August 2026. Completion rate is calculated by dividing submissions by form starts.