kondeor.ai · Blog

← All posts

Researchers Are Learning Differently, Not Less: AI in Data Analysis

15 November 2026

I've been wanting to write this blog post for a long time. Not so much because researchers have been asking questions about it, but because kondeor.ai has sparked intense discussions about which competencies should be taught in academic training — outside of dedicated statistics or market research programs — in the area of data analysis. A LinkedIn post by Harald Dragan about "The wheel that never stops spinning: When AI analyzes our data" finally prompted me to do it.

Over the more than 30 years I've spent working with quantitative data analysis, the available software tools have evolved step by step. Interfaces became easier to use, the range of methods grew larger. But regardless of the program used, you always had to be familiar with the fundamentals: which statistical test to use when, where to find it in the software, and how to read the results. These very questions had kept some people from running their own ad hoc analyses on data we had collected or systematically recorded together in joint projects. To give them a tool that removes this barrier, I began developing kondeor.ai. Very quickly, it became clear that it was also radically changing my applied quantitative data analysis courses at universities.

kondeor.ai relieves researchers of many decisions around the fundamentals mentioned above. But that does not mean they lose the skills that are essential for quality results.

Does someone running do-it-yourself market research for a company — or writing an academic thesis in a non-specialist field — really need to know how to test the normality of two groups for a t-test? Or whether they need the Levene test, which is controversially discussed in the literature, and if so, what consequences its result has? Or which effect size measure to calculate in that case? (These considerations naturally do not apply to professional market research institutes, from which these competencies will rightly continue to be expected.) Isn't it far more important to identify the relevant issues beforehand and formulate a research question — and then draw the right conclusions from the analysis results in answering that question?

That said, there are a few points that cannot be delegated to a data analysis program in general, or to kondeor.ai in particular. These are areas where you need to be competent and take an active role yourself. If you follow this, then the "AI wheel of death" described and depicted by Harald Dragan breaks down at precisely these points:

Even before the first point in the wheel — "AI checks the data" — important decisions must be made in the content, research, survey, and sampling design that AI cannot take off researchers' hands. In the concrete implementation, however, it can contribute valuable support. Just to give one example: kondeor.ai can provide guidance on academically correct formulation of research questions. When planning the content and survey design, decisions about the measurement levels of variables must be made. Even AI cannot correct mistakes at this stage during data analysis. If in doubt, kondeor.ai can ask about the measurement level; this is another area where prior engagement pays off. And as part of the content design, the hypotheses to be analyzed should be formulated. Only then can significance values be read confirmatorily, and the results do not have to be treated primarily as exploratory. Where necessary, kondeor.ai will point this out. Important: Parts of the data checking in kondeor.ai work without AI as well.

Before "AI selects and runs the test" (the second point in the wheel), researchers must themselves decide, for example, how to handle missing values, or whether skewed distributions in the data should be adjusted. kondeor.ai will ask about this when relevant, but does not decide autonomously. Important: Test selection in kondeor.ai also works without AI; analysis with R does anyway.

On the point "AI writes the report": ideally, researchers can use the information provided by kondeor.ai to judge whether a chosen method is appropriate, whether its results contribute to answering the original question, and whether it should therefore be included in the report. For instance, kondeor.ai draws attention to insufficient statistical power.

"The client has AI summarize it" does not apply to reports that you create yourself in a business or academic context using kondeor.ai. When AI generates a summary across all analyses, it must be reconciled with the researcher's expert knowledge. A binding executive summary of this kind — especially when the report goes to third parties — should also include the limitations of the findings. And that makes a further summary by yet another AI unnecessary.

"Decisions are made faster" and "Researchers learn less" are points five and six of the wheel presented by Harald Dragan. But these are not action fields — they are outcome fields. I would rephrase the first statement as "Analysis processes move faster."

That leaves the central claim of the wheel: "Researchers learn less." This is only true if kondeor.ai is used (improperly) as an autonomous agent. But kondeor.ai is not an autonomous agent — it is a tool for researchers. If they already have the competencies described above, or if these are taught to them in courses, then learning shifts. "How do I have to do it?" moves into the background; "What can I do?" (in terms of analysis and use of results) comes to the fore. Experience so far in companies and in university courses shows that, freed from the occasional statistical "ballast," people engage more deeply with the purpose and meaning of data analysis.

In one course project, for example, students not only developed detailed data-driven marketing scenarios for an organization. They also transformed the initial hypothesis model into a model of central insights — and in doing so, without being asked, made the leap from mechanically working through hypotheses to a meaningful interpretation of the results.

Researchers are therefore learning differently, not less. And in the end, that can actually amount to more — and be more valuable.

To kondeor.ai →