Habr AI: LLMs Can Take Over Routine in Business Research—But Not Strategy

A comprehensive analysis of the future of product and marketing research boils down to a simple idea: LLMs are not equally useful for all types of research work. The more rigidly a framework is established in advance — what to measure, how to ask, and how to interpret the answer — the easier it is to automate this work.

Three Levels of Research

The author proposes viewing business research not as a set of separate methods, but as work with a "knowledge matrix" — a system of distinctions through which a company describes the market, user, and product. At the first level, this matrix is already ready: the researcher simply fills it with data. At the second level, the framework is refined as the work progresses: categories change, segments are recombined, and the user behavior model gradually adjusts to reality.

At the third level, the framework itself becomes the subject of analysis — how the company defines value, problem, loyalty, or success in the first place. From this logic follows the text's main conclusion: the question is not whether LLMs can conduct research in general, but at what level the specific task sits. If a business needs to measure predetermined parameters, the model will succeed much sooner.

If the team needs to reconsider the very concepts through which it describes the market, current language models are insufficient. This is why the debate over completely replacing researchers here seems too crude: different classes of tasks will be automated at different speeds.

Where LLMs Are Already Strong

The most obvious candidate for automation is first-level research. Here the company already has ready-made metrics, question templates, and interpretation rules. In essence, this is not about finding new research logic, but about the rapid execution of a formalized procedure. That's why the author believes many such tasks could have been automated earlier, and LLMs simply drastically reduce cost and entry barriers.

• sales funnels, NPS, CSI and other metrics with fixed calculation rules
• one-off A/B tests for comparing pre-defined variants
• pricing research like Van Westendorp, Gabor-Granger, and conjoint approaches
• structured CustDev interviews and usability tests with rigid scenarios
• feature prioritization through models like Kano, MaxDiff, and TURF

"LLMs don't create a new class of capabilities, but only remove the

costs of formalization."

With more complex second-level tasks, the situation is more nuanced. Here a simple prompt or RAG is insufficient: the model must not only process responses but also gradually refine the very set of distinctions on which the analysis is built. The article names LoRA and Representation Engineering among suitable approaches — methods that change model weights or activations, thereby allowing the adjustment of its semantic field. In other words, the author suggests that LLMs may help with segmentation of complex audiences, development of decision-making models, and refinement of research categories, but this is no longer "chat with documents," but deeper system tuning.

Where the Limit Lies

The main limitation begins at the third level, where research must not fill or refine an existing framework, but disassemble it and reassemble it anew. These are tasks where the team asks not "why is NPS falling" but "what exactly do we call loyalty and why do we consider it important." This also includes research into brand language, cultural codes, organizational discourses, and strategic concepts through which the company sees its problems and opportunities.

According to the author, the current architecture of LLMs hits a fundamental limit here. The model can generate interpretations, debate itself in multi-agent setups, and even use self-reflection, but all of this remains work within the same system of distinctions. Such a loop can improve the answer, but does not transform the model itself into an object of sustainable transformation.

Therefore, it can support the researcher, suggest moves, and accelerate analysis, but cannot replace the human where it's necessary to reassemble the research perspective itself.

What This Means

The practical conclusion is harsh: most of what is today called product and marketing research will become an automatable service. The value of people will shift to where it's not enough to simply count, compare, and code responses, but to change the language of problem framing, notice hidden frameworks, and connect business, culture, and strategy into one research picture. From this comes the author's forecast: instead of rigidly separated roles like UX researcher, CX manager, or marketing analyst, there will grow demand for curator-researchers who know how to manage an ensemble of AI tools.