The board meeting begins. Before each participant lies an analysis created by the AI system: 47 data points, five scenarios, three prioritized recommendations for action. The system’s recommendation is clear. The discussion is also clear. After twenty minutes, the decision is made. And no one asks whether what the algorithm suggested was actually the right decision.
AI can improve decisions. It can also have the opposite effect if accepted unchecked. The difference lies not in the algorithm, but in the manager who interprets its results.
A manager whom I advised had introduced an AI-powered forecasting system for his division heads. After six months, he observed a pattern: decisions were made faster, discussions were shorter, and certainty increased. At the same time, strategic discussions became superficial because the system’s results suggested a clarity that had not been earned through independent thought. “We decide faster and feel more secure. But we question less,” he summarized. This was not an anti-AI stance. It was an honest observation about the price that even useful tools demand.
There is a reflexive polarization on the topic of AI in management: some expect a revolution in decision quality, while others warn of the loss of human judgment. Both perspectives fall short. AI changes how managers process information and evaluate options. But it does not replace the responsibility for what an organization does and why. Three levers help to use the technology correctly.
Daniel Kahneman described two thinking systems in his research: System 1 works fast, intuitively, and automatically. System 2 works slowly, analytically, and consciously. Both have their strengths and weaknesses. AI is an amplifier of both systems, but it does not work equally well everywhere.
Where AI actually creates value: in processing large amounts of data that exceed human analytical capabilities. In identifying patterns that we overlook with the naked eye. In simulating scenarios whose number exceeds human imagination. In reducing cognitive biases by consistently applying the same criteria where humans would decide differently depending on their daily mood.
Where AI does not help: with decisions where the data basis is thin or one-sided. In situations that have not occurred in the past. In evaluations that have ethical, cultural, or political dimensions that cannot be algorithmically captured. And everywhere where the assumptions on which the system was trained no longer match current reality.
| Where AI Creates Value | Where Management Remains Essential |
|---|---|
| Analyze large data volumes | Evaluate whether the data answers the right question |
| Recognize patterns | Assess whether the pattern is causal or random |
| Simulate scenarios | Decide which scenario is desired |
| Reduce cognitive biases | Set values and priorities not embedded in the system |
| Derive recommendations consistently | Recognize the system’s blind spot |
The most common misjudgment: AI is treated like an oracle that provides answers, instead of a sparring partner that tests assumptions. In psychology, this phenomenon is known as automation bias: the tendency to trust automated suggestions and discard one’s own contradictory observations in favor of the algorithmic recommendation, without anyone noticing this process. Those who consult algorithms instead of questioning them delegate responsibility to a system that cannot bear this responsibility.
Every AI system has blind spots. They lie in the data it was trained on, in the assumptions of its developers, and in the limits of its model architecture. Anyone who works with AI without knowing these blind spots trusts a system whose limitations they do not understand. AI systems inherently look in the rearview mirror. They extrapolate the past. If the training data is biased or outdated, the principle of computer science applies: “Garbage in, garbage out.” An algorithm optimizes what exists, but it cannot foresee developments that have never occurred before.
A division head whom I supported in introducing an AI-powered procurement system experienced this in a critical situation: the system recommended delisting a long-standing secondary supplier in favor of the primary provider. The reasoning was logical: better price, comparable delivery reliability in history. She followed the recommendation. When the remaining main supplier ran into a supply crisis months later, there was no alternative. The system had calculated correctly but had not seen a risk that was not in any historical data: the dependency created by the elimination of the second source. The lesson she learned from this: “Since then, with every recommendation, I first ask myself what risk it cannot account for.”
Most AI systems are a black box internally. No one in the company can understand in detail how a recommendation came about. Therefore, the question “How does the system work?” rarely leads anywhere. What a manager can assess, however, is the recommendation in light of their own knowledge. Three questions help with this. First: Does the recommendation fit what I know about the context, or does it contradict my experience? A contradiction is not proof that the AI is wrong, but a signal to look more closely. Second: What can the system structurally not know? Events that have never occurred, soft factors like relationships and trust, strategic intentions that are not in any data. Third: What would have to be true for this recommendation to be wrong, and how likely is that? This question transforms a seemingly clear recommendation back into what it is: a hypothesis under assumptions. Thus, critical examination also becomes a question of risk management: those who do not consider the blind spots cannot assess the risk of the recommendation.
In my consulting practice, I see this question about assumptions asked far too rarely. The slide with the recommendations seems convincing. No one wants to be the one who delays the meeting with the question “but are the assumptions correct?” This hesitation is understandable, but it is dangerous. The critical examination of assumptions is not a delay. It is the prerequisite for a responsible decision.
The third lever is the most uncomfortable because it demands discipline. Consistently separate the preparatory work, where AI can help, from the decision itself, which you as a manager must make.
The preparatory work includes analysis, scenario building, and evaluation of options based on defined criteria. Here, AI can improve speed and depth. The decision itself is different. It involves weighing values that are not embedded in the system: What responsibility do we bear towards employees? What risks are we willing to take? What kind of company do we want to be? These questions cannot be delegated, neither to employees nor to algorithms.
In practice, this means: Treat AI recommendations as input, not as the final result. Use them to test your own assumptions, not to replace them. And create a decision-making culture where the question “Is that really true?” is not seen as distrust of the system, but as an expression of professional responsibility. Those who accept AI recommendations unchecked create a new form of information bubble: one in which no one contradicts anymore because the algorithm has already dictated the consensus.
Keep your judgment sharp. Judgment develops through practice. Those who permanently delegate their own judgment to systems will lose the ability over time. An organization whose managers can no longer fully comprehend their own decisions is vulnerable precisely when the system gives a recommendation that needs correction. This applies not only to strategic directions but also to operational control, where AI increasingly takes over routine decisions. And the opposite is just as dangerous: those who do not act at all out of distrust of AI run the risk of inaction.
First: For which AI-supported recommendation did you last actively question the underlying assumptions? If you cannot think of a specific example, consistently question your system’s next recommendation.
Second: Can you identify, for the most important systems in your area of responsibility, what data they were trained on and what assumptions they contain? If not, this knowledge is a prerequisite for responsible use.
Third: Are AI recommendations discussed as input or accepted as the final result in your management meetings? Observe consciously next time and ask about the assumptions, even if no one else does.
AI changes the conditions under which managers make decisions, but it does not change the responsibility for what is decided. The temptation to delegate this responsibility to a system is great because it promises relief. But it does not eliminate the demand on management; it merely shifts it to where no one perceives it.
AI takes work off managers’ hands, but not responsibility. Those who confuse this delegate their judgment to a system that has no judgment, but only calculates. Use the recommendation for what it is: a well-informed starting point. The decision remains yours.
Decisions Under Uncertainty – How to remain capable of action with incomplete information, with or without algorithmic support.
Technical Literacy for Managers – Why you don’t need to master the technology yourself, but understand what it can and cannot do.
All Insights can be found in the overview.
Proven tools and models for self-application are available under Solutions.
If you want to take these thoughts further for your company, a no-obligation initial conversation is worthwhile.
