Which qualitative data analysis software should you use?

Direct answer: The open evidence does not crown a winner. What it does measure is learning: Salmona et al. (2022) surveyed 355 researchers in 29 countries, and 73% said they learned by trial and error or taught themselves. Pick the program you can learn with support, not the one with the longest feature list.

A supervisor's usual question is "why this software and not that one?", and the honest answer starts with who is asking. Salmona et al. (2022) collected their 355 responses from 29 countries through listservs, with 226 respondents in the United States, 64 in Australasia and 26 in Europe, and found that only 21% had been trained through an academic course. Their full text is open at Bowling Green State University, and they conclude that "technological confidence should not be confused with methodological competence" (Salmona et al., 2022).

For a student the practical reading is that the program matters less than the support around it. NVivo, ATLAS.ti and MAXQDA all store transcripts, hold codes and retrieve tagged passages, and none of the open studies below tested one against another. Zamawe (2015) reflects on NVivo alone, so his account cannot show that NVivo beats its rivals. A university licence, a workshop at your library and a supervisor who already uses the program are therefore stronger reasons to choose than any comparison chart.

If your design is still undecided, settle it first. How to do a thematic analysis describes the coding steps the program will later hold, and qualitative research methods compared explains what interviews, focus groups and case studies give you to code.

Direct answer: It stores your sources, holds your codes and retrieves every passage tagged with one idea. It does not decide what the data mean. Zamawe (2015) found that a retrieval question taking over four hours by hand took seconds in NVivo, yet the analysis stayed his own.

The reflection by Zamawe (2015) is a single researcher's account, based on ten interview files, so treat the figures as an illustration rather than a benchmark. Even so, it is one of the few open papers that attaches times to tasks. Asked how many women with no child had completed all antenatal visits, manual searching took over four hours, while the program answered in seconds. Combining two coding nodes and adding two more took less than two minutes. He adds that with more than 1,000 transcripts, reading every source in detail by hand would be very unlikely.

The same paper draws the limit. Zamawe writes that "unlike statistical software, the main function of CAQDAS is not to analyse data but rather to aid the analysis process, which the researcher must always remain in control of" (Zamawe, 2015, p. 13). CAQDAS stands for computer-assisted qualitative data analysis software, the family that holds the three programs above. He also admits he chose the software before the design, and he found the program hard to learn despite advanced computer skills.

Source: Zamawe (2015), https://doi.org/10.4314/mmj.v27i1.4, read on 31 October 2026.

Direct answer: No. Open tests of GPT-style models against human coders show agreement that ranges from weak to substantial. Kappa was 0.401 on 20 interviews (Li et al., 2024) and 0.61 to 0.65 for multi-label themes (Wen et al., 2026), so AI output is a draft you must check.

Several programs now advertise AI-assisted coding, and the open studies below tested general chatbots rather than those built-in features, so they show what such models can do, not how any vendor's button performs. Cohen's kappa is the agreement statistic they report: 0 means chance-level agreement and 1 means perfect agreement.

Source: the three open-access papers named in the table, read on 31 October 2026.

Prescott et al. (2024) also timed the work: the AI analysis took 20 minutes (SD 3.5) against 567 minutes (SD 106.5) for human coders, and the authors write that "the relatively lower reliability in coding between them suggests that hybrid approaches are necessary" (Prescott et al., 2024). Wen et al. (2026) found GPT-4o strong on sentiment (κ = 0.91 to 0.95) but near zero on judging evidence of impact (κ ≤ 0.01), a task that depends on context. The pattern is consistent: models are quick and decent at sorting, and weak where interpretation is needed. Whether your institution allows external processing of interview data is a separate question, covered in which AI tools are allowed in university work.

Direct answer: Run a small pilot. Code one transcript in the program and again by hand, compare the two, and check your ethics approval before switching on any cloud or AI feature. A pilot costs an afternoon, while discovering a mismatch after coding a full dataset costs weeks.

The evidence gives three reasons for a pilot. First, learning is the common bottleneck, so you want to find out early whether you can work in the interface. Second, Zamawe (2015) warns that choosing the tool first can bend the design to fit it, so ask whether your research question would look the same without the program. Third, agreement with an AI draft varied widely, and Prescott et al. (2024) reported code agreement as low as 36%, so a side-by-side check on one transcript is proportionate rather than fussy.

Practical order: ask your supervisor which program your department licenses and which they accept; list the data you will code, such as transcripts, images or audio; read your participant information sheet and ethics approval to see where data may be stored; then run the pilot. If the pilot shows the program saves time on retrieval but not on thinking, that matches the published pattern and is a reason to continue.

Record the program, its version and the date in your methods chapter. Examiners tend to ask how the software was used, and a plain sentence that you coded and interpreted the data yourself, using the program for storage and retrieval, answers it. For the interview design that feeds the program, see semi-structured interviews for a dissertation.

Is NVivo better than ATLAS.ti or MAXQDA?
The open papers do not rank them. Zamawe (2015) writes about NVivo alone, so nothing here shows a winner. Choose by your licence, your data types and the help your department offers.

Is free qualitative data analysis software good enough?
Free options exist, but the studies read here did not test them against paid programs. Ask your supervisor whether a free tool is acceptable before you code your whole dataset in it.

Can the software analyse the data for me?
It sorts and retrieves; you interpret. Zamawe (2015) says the researcher must always remain in control of the analysis, and the AI tests above agree with human coders only partly.

Does AI coding save time?
Yes, in the open study that timed AI analysis: 20 minutes against 567 (Prescott et al., 2024). The agreement figures show you still need time to check the output.

Do I have to name the program in my thesis?
Yes, name the program, its version and what you used it for. Methods sections are read for exactly that detail.

If you are unsure how to set up your codes or how to report the program in your methods chapter, you can book a short consultation on your coding setup through our SPSS and data analysis support. The coding, interpretation and writing remain yours, and the service does not set a grade target.

• How do you do a thematic analysis?: the coding method your program holds
• Which AI tools are allowed in university work?: finding your institution's rule
• Which AI and research tools help at each stage of a project?: the wider workflow

• Li, K. D., Fernandez, A. M., Schwartz, R., Rios, N., Carlisle, M. N., Amend, G. M., Patel, H. V., & Breyer, B. N. (2024). Comparing GPT-4 and human researchers in health care data analysis: Qualitative description study. Journal of Medical Internet Research, 26, Article e56500. https://doi.org/10.2196/56500
• Prescott, M. R., Yeager, S., Ham, L., Rivera Saldana, C. D., Serrano, V., Narez, J., Paltin, D., Delgado, J., Moore, D. J., & Montoya, J. (2024). Comparing the efficacy and efficiency of human and generative AI: Qualitative thematic analyses. JMIR AI, 3, Article e54482. https://doi.org/10.2196/54482
• Salmona, M., Kaczynski, D., & Lieber, E. (2022). Improving research practices through user insights of qualitative and mixed methods data analysis technology. Mid-Western Educational Researcher, 34(3), Article 3. https://scholarworks.bgsu.edu/mwer/vol34/iss3/3
• Wen, C., Clough, P., Paton, R., & Middleton, R. (2026). Leveraging large language models for thematic analysis: A case study in the charity sector. AI & Society, 41(1), 731-748. https://doi.org/10.1007/s00146-025-02487-4
• Zamawe, F. C. (2015). The implication of using NVivo software in qualitative data analysis: Evidence-based reflections. Malawi Medical Journal, 27(1), 13-15. https://doi.org/10.4314/mmj.v27i1.4

This article is part of the MAAS Journal series for students. MAAS is an academic support partner: experts guide students through each stage of their own work, and we do not complete assessed work on a student's behalf.