Sber study: how AI quietly substitutes its own assumptions for strategic decisions

Top executives consult AI more often than other groups: according to Gallup 2025, this is now a stable trend. Managers let models into strategic tasks such as forecasting market development, assessing risks, setting priorities and choosing a direction for growth. And here lies the root problem: it is very hard to formulate a strategic question well. It usually retains uncertainty, contradictory data, implicit company and industry context, and vague terms.

Where AI becomes dangerous

Missing context and internal contradictions in how a task is framed do not make the model stop and ask for clarification. Instead, it fills the gaps with its own assumptions. And it does not say that it is doing so. The answer looks confident, logical and full of reasoning, but it rests on distorted or entirely invented data.

This is not theory: it has been confirmed experimentally. Sber's Laboratory of Neuroscience and Human Behavior studies psychology, cognitive processes and how thinking changes in the age of AI. The team conducts research at the intersection of cognitive science, business and AI. Based on the results, it prepares Think Tank materials for top executives: analyses of how new technologies change decision-making and the work of organizations.

What AI substitutes

When a model meets a gap in information, it silently fills it with four types of assumptions:

• Lack of context: the model picks the most likely scenario from its training data, ignoring the uniqueness of your situation and company
• Contradictory data: it prefers one interpretation and hides the conflict between sources instead of pointing it out
• Implicit context: it fills in gaps in specialized knowledge about your industry, competitors, legislation and company culture
• Vague terms: it interprets "strategic priority", "risk" or "competitive advantage" by its own understanding, not by your company's definition

The executive gets an answer that sounds well argued, with a logical chain of reasoning, but contains hidden errors. The trouble is that it is practically impossible to spot them while reading. They are buried in the very texture of the answer.

Strategy: costly errors surface late

In operational decisions, a model's error shows up quickly. A wrong short-term forecast is exposed the following month. A strategic error, by contrast, surfaces months or quarters later, when the company has already spent substantial resources, hired people and changed plans. And because the AI's advice looked confident and logical, the executive may not have sought alternative opinions, outside expertise or contradictory data.

What this means

AI is good for quick analysis of facts and for producing a first draft. But in strategy it is dangerous as an arbiter and decision adviser. It should be used as an assistant: for rapid rough analysis, finding additional arguments and generating alternatives. The final decision stays with the human, and above all, executives need to become far more critical of the model's advice.