Can AI Help With Qualitative Data Analysis? What Works and What Doesn't

TL;DR:

• Yes, AI can help with the slow mechanical parts of qualitative data analysis: draft transcripts, summaries, first-pass coding against your own codebook, and finding passages you missed.
• It cannot interpret for you. In reflexive thematic analysis (Braun and Clarke), your own reading is the method, and in 2025 a letter for 419 qualitative researchers, Braun and Clarke among the authors, rejected generative AI for this work.
• NVivo, ATLAS.ti and MAXQDA all have AI features as of November 2026. Most send text to a cloud service; MAXQDA's Premium tier and open-source Whisper can run on your own machine.
• Under Kenya's Data Protection Act 2019, you must be able to prove consent, and sensitive data such as health information may leave Kenya only with consent and proper safeguards. Pseudonymise transcripts before any cloud tool sees them.
• Check every output against the transcript, keep a reflexive journal, and state exactly what the tool did in your methods section.

Who this guide is for

This guide is for postgraduate students writing a thesis from interviews, focus group discussions (FGDs) or open-ended answers, and for NGO monitoring and evaluation (M&E) staff working with key informant interviews (KIIs) and FGDs. You do not need any technical background.

It covers words, not numbers. For survey statistics, read our guide on using AI for data analysis in postgraduate research. For indicator tables and donor narratives, read how NGOs can use AI for M&E data analysis.

The short answer: where AI helps and where it does not

AI is useful for tasks you can check line by line against the transcript. It is weak, or wrong for the method, wherever the task is interpretation.

First, know your method

Your method decides how much AI you can defend, so settle it before choosing a tool.

Reflexive thematic analysis (Braun and Clarke)

Virginia Braun and Victoria Clarke now call their widely cited approach "reflexive TA". On their own website they set out six phases:

1. Familiarising yourself with the dataset
2. Coding
3. Generating initial themes
4. Developing and reviewing themes
5. Refining, defining and naming themes
6. Writing up

They describe qualitative analysis as "a skilled, situated, subjective process". Your background, reading and judgement are part of the method.

That is why a 2025 open letter matters. In Qualitative Inquiry, Jowsey, Braun, Clarke, Lupton and Fine, writing for 419 experienced qualitative researchers from 32 countries, argued that generative AI does not fit reflexive approaches, because these are human research practices. They also rejected it on social and environmental justice grounds.

If your proposal says "reflexive thematic analysis", your examiners may know this letter. You have three honest options:

• Analyse without AI, and use it only for admin tasks such as tidying transcripts.
• Use AI in narrow, checkable ways (such as a second search for passages you missed) and justify it in your methods chapter.
• With your supervisor's agreement, switch to a method that suits structured coding.

In our view, the first is the safest for a reflexive TA thesis, and the second can work if you are open about it. Agree it with your supervisor before you start.

Approaches with a fixed codebook

Framework analysis, directed content analysis and much evaluation coding start from a codebook agreed in advance. AI fits better here, because the task is closer to applying clear rules: "Does this passage match code 3, as defined?" You can measure how often the AI agrees with you. The codebook, final decisions and conclusions still stay yours.

What NVivo, ATLAS.ti and MAXQDA offer (as of November 2026)

All three major packages now include AI features, and each has real strengths. We do not recommend one: use what your institution licenses and your supervisor knows. Features below come from each company's own help pages, checked on 11 November 2026. Privacy statements are the vendors' own claims.

Details that matter in practice:

• NVivo: In Suggest Child Codes, every suggestion is ticked by default. Untick the ones you do not want before you click OK.
• ATLAS.ti: Both AI coding features work per paragraph, so put each question and speaker on its own line. PDFs give poor results. The manual warns that output can be incorrect and "may encode social biases".
• MAXQDA: AI Assist "is designed to support your analysis, not replace your own interpretation", says its help page. The free version allows limited prompts and single-document coding; Premium adds more prompts, multi-document coding and local AI.

Transcription: your options

Transcription is where most researchers first meet AI, and where errors slip in quietly. Three routes:

• Transcribe in Microsoft Word. Microsoft 365 subscribers can transcribe up to 300 minutes of uploaded audio a month (30,000 with a Copilot licence), says Microsoft. It labels speakers ("Speaker 1", "Speaker 2") and saves recordings to OneDrive. As of 11 November 2026, Swahili is not on its language list.
• Whisper, on your own computer. OpenAI's Whisper is an open-source speech recognition model under the MIT licence. Swahili is in its language list, though its documentation warns that performance "varies widely depending on the language". Your audio stays on your machine. Our guide to running open-source AI models locally explains the hardware you need.
• Built into your analysis software. ATLAS.ti has had auto transcription since version 25.0.0 (December 2024). MAXQDA offers MAXQDA Transcription, now a separate add-on, which its help page says runs on servers in the Netherlands and Germany. These run in the cloud, so the same consent questions apply.

Kenyan interviews often mix English, Kiswahili, Sheng and a mother tongue in one answer. Test any tool on a short recording made with a colleague's consent first. Our postgraduate data analysis guide explains why transcription errors, including invented sentences, can be serious.

Read every AI transcript while listening to the audio. Correcting it is also how Phase 1, familiarisation, begins.

Consent, anonymising and Kenya's Data Protection Act 2019

Interview data is personal data, often sensitive. Kenyan law and your ethics approval apply before any AI tool touches it. This is a summary, not legal advice.

Ethics approval and the NACOSTI licence

NACOSTI's licensing guidelines say research involving human subjects needs approval from a NACOSTI-accredited Institutional Ethics Review Committee before you apply for a research licence. Your approved protocol and consent form describe how data will be handled. If they say nothing about cloud transcription or AI tools, ask your ethics committee first. Silence is not permission.

What the Act says

These sections of the Data Protection Act 2019 matter most for qualitative researchers:

For more, see whether staff can use ChatGPT under Kenya's Data Protection Act and, for beneficiary stories, data privacy for NGOs using AI. The Data Protection Commissioner's Guidelines page also lists a guidance note on processing for research purposes.

How to anonymise a transcript before AI sees it

Removing names is not enough. In a small community, a job title or a story can identify someone as clearly as a name.

1. Work on a copy. Keep the original in a secure, password-protected place.
2. Build a key. In a separate file, list each real detail and its replacement: "Participant 07", "a sub-county hospital", "a Nairobi estate".
3. Replace direct identifiers: names, phone and ID numbers, emails, addresses.
4. Replace indirect identifiers: village, estate and school names, church and chama names, clan names, unusual job titles and rare events ("the only woman MCA in the ward").
5. Generalise: "a mother of four in her thirties", not exact ages and children's names.
6. Read it as an outsider. Could a neighbour or employer recognise this person? If yes, generalise further.
7. Store the key apart, and never upload it.

A local model can flag names you missed, but a person makes the final call. If a transcript cannot be safely pseudonymised, keep it out of cloud AI tools.

Keeping yourself in the analysis: reflexivity

Reflexivity means noticing how you shape the analysis: your background, assumptions and relationship with participants. AI adds a new influence to notice. Keep a reflexive journal, and when you use AI, note:

• What did I ask the tool to do, and why?
• What did I accept, change or reject, and did any suggestion pull my thinking somewhere new?
• What did it miss: silences, contradictions, humour, tone, cultural meaning?

Watch for generic code labels ("challenges", "support", "communication") that could fit almost any dataset. When a suggestion feels obvious, go back to the transcript and ask what is specific to your participants. The ATLAS.ti warning about social bias applies to every tool.

A safe workflow in eight steps

1. Agree the plan with your supervisor or M&E lead: which tasks, which tool, how you will check it.
2. Check consent and ethics approval cover the tool, especially if data leaves Kenya.
3. Transcribe and correct every transcript against the audio, locally if you can.
4. Pseudonymise, keeping the key separate.
5. Read everything yourself and make early notes before any AI summary exists.
6. Code a few full transcripts by hand and draft your codebook. Only then, if your method allows, use AI to test it or as a second pass.
7. Check every AI output in context, and record how many suggestions you accepted and rejected.
8. Develop themes and write up yourself. Use AI, if at all, to challenge your draft. Disclose exactly what it did: the tool and version, each task, how you checked it, what it did not do, and how data was protected.

Prompts you can copy

Use these only with pseudonymised text, in a tool your ethics approval covers. Replace the square brackets.

1. Check your anonymisation

Below is a pseudonymised interview transcript from a study in Kenya. Do not summarise or analyse it. List every word or phrase that could still identify anyone, directly or indirectly: names, places, institutions, job titles, dates, unusual events. Quote each one, say why, and suggest a more general replacement. Transcript: [paste]

Run this in a local model where you can, since the text may still hold identifiers.

2. First-pass coding with evidence (codebook approaches only)

Apply only the codes in my codebook below to this transcript. Do not invent new codes. For each coded segment, give the code name, the exact quote, and one sentence explaining why it fits the definition. If a passage could fit two codes, say so. If you are unsure, mark it [UNSURE]. Do not summarise the interview or suggest themes. Codebook: [paste] Transcript: [paste]

The exact quote lets you check each decision quickly. Expect to reject some of them.

3. Look for what you might have missed

I have coded this transcript for the code "[code name]", defined as "[definition]". Here are the passages I coded: [paste]. Here is the full transcript: [paste]. Find any other passages that might fit this definition, and any passages that contradict or complicate it. Quote each one exactly. Do not tell me whether my coding is right.

Contradictions are often where the interesting analysis lives.

4. Challenge your draft theme

Here is a draft theme from my analysis, with its definition and three supporting extracts. Act as a critical examiner. Ask me five hard questions about whether this theme is well supported, whether it is a theme or just a topic summary, and what alternative readings of the extracts exist. Do not rewrite the theme. Theme: [paste]

5. Draft your disclosure statement

I used [tool and version] in my qualitative analysis for these tasks only: [list]. I checked the output by [method]. I did not use it for [list]. Draft a short, factual paragraph for my methods chapter that states this clearly, in the first person, without overstating or understating the AI's role.

Edit the draft yourself; you are responsible for every word. Our post on using AI in postgraduate research without crossing the integrity line covers disclosure and university rules in more depth.

What goes wrong, and how to avoid it

Questions people ask

Can ChatGPT do thematic analysis for me?

It can produce something that looks like one, but the analysis would not be yours. In reflexive thematic analysis your interpretation is the method, and Braun, Clarke and hundreds of other qualitative researchers reject generative AI for it. Use AI, if at all, for narrow tasks you can check.

Is AI-assisted coding acceptable in a Kenyan university thesis?

It depends on your university's rules, your supervisor and your method, and rules differ between institutions. Agree it in writing early, and describe it fully in your methods chapter.

Which is better for AI features: NVivo, ATLAS.ti or MAXQDA?

Each has strengths. In our reading of their documentation, NVivo's AI Assistant focuses on summaries and child code suggestions, ATLAS.ti offers the most automated coding (still in beta), and MAXQDA explains each coding suggestion and offers local AI on its Premium tier. Use what your institution licenses, and test it on pseudonymised data.

Will AI transcription work for Kiswahili and Sheng?

Results vary by tool and speaker. Microsoft Word's Transcribe does not list Swahili as of November 2026; Whisper includes it, with a warning that quality varies by language. Test a short consented sample and correct every transcript against the audio.

Can M&E teams use AI to code KIIs and FGDs?

Yes, with a clear codebook, pseudonymised transcripts and a human check of every coded segment. Agree the approach with the evaluation commissioner, and say in the report what the tool did.

Sources

All sources last checked 11 November 2026.

• Braun and Clarke, Thematic Analysis website (University of Auckland and UWE)
• Braun and Clarke, Doing reflexive TA (the six phases)
• Jowsey, Braun, Clarke, Lupton and Fine (2025), "We reject the use of generative artificial intelligence for reflexive qualitative research", Qualitative Inquiry, research record at Bond University
• NVivo 15 help, What's new (release notes)
• NVivo 15 help, AI Assistant
• NVivo 15 help, AI Assistant: Suggest Child Codes
• ATLAS.ti 26 manual, AI Coding
• ATLAS.ti 26 manual, Intentional AI Coding
• ATLAS.ti, Update history
• ATLAS.ti, AI data security and privacy
• MAXQDA, AI Assist product page
• MAXQDA help, What is AI Assist?
• MAXQDA help, Automatic transcription
• Microsoft Support, Transcribe your recordings
• OpenAI, Whisper on GitHub
• OpenAI, Whisper language list (tokenizer.py)
• The Data Protection Act, No. 24 of 2019 (copy published by the ODPC)
• Office of the Data Protection Commissioner, Guidelines page
• NACOSTI, Guidelines on Research Licensing and Institutional Affiliation

Where FuKazee fits in

FuKazee trains postgraduates, researchers and NGO teams to use AI carefully in real work. We assume nobody knows AI yet, avoid jargon, and show the free version of each tool. Nobody becomes an expert in a few days, so we focus on safe habits: protecting participants' data, checking every output, and staying the author of your analysis. FuKazee has trained more than 3,000 people.

For Masters and PhD students, the ScholarShift Accelerator is a four-week live programme on using AI in research without compromising academic integrity: AI ethics and disclosure, prompting, literature reviews and academic writing. For NGO and M&E teams, we build sessions around your own reports and data: see AI training for NGOs in Kenya.

To ask about training for your research group, department or M&E team, email us at lead@fukazee.com