The silent conductor in decision mechanics: How recommender AI shapes our behaviour

When a decision-maker in a modern company faces a choice, for example a piece of software, a service provider or a partner, it has long since stopped being only about product features. On digital platforms, in online markets and in SaaS ecosystems, the way options are presented, weighted or prioritised is increasingly part of the decision itself. Algorithmic systems, such as recommender systems (“Recommender Systems”, RS) or AI agents with autonomy, sit quietly at the table too. They do not always influence visibly, but they often influence effectively which decisions get made.

For entrepreneurs this means: anyone who designs innovation and business processes must factor in the design of the decision architecture. This analysis engages intensively with the difference in how recommender systems and algorithmic decision autonomy influence the behaviour of users and customers, and with what can be derived from this for entrepreneurial action.

Theoretical background

Two central concepts stand at the centre.

Recommender systems (Recommender Systems, RS)

An RS is a system that, out of a large set of options, proposes a narrower subset of possible decisions to users, based on user behaviour, item properties, similarities and so on. According to Dietmar Jannach (Universität Klagenfurt), such systems influence ”[…] independent of such specific goals […] the users’ choices and their behavior as a matter of principle.” They therefore act not only as a technical function but as a structural specification of what is perceived in the first place.

Algorithmic decision autonomy

It describes how strongly an AI system intervenes in a decision process, that is, how much scope for action the machine takes over and how much freedom of decision remains with the human. In simpler terms: the algorithm can give a recommendation, co-decide, or make the decision entirely on its own.

Depending on the degree of this autonomy, the way users react to the suggestions changes. An empirical study by Yuejiao Fan and Xianggang Liu (2022) examined this relationship experimentally. Test participants were confronted with differently “autonomous” recommender systems:

• once with low autonomy (the system only makes suggestions),
• once with medium autonomy (the system recommends and actively rates),
• once with high autonomy (the system selects a product directly for the participants).

The result was remarkable: the effect of algorithmic autonomy on the purchase decision does not follow a straight line but an inverted U-shaped curve (“inverted U-curve”). This means:

• Too little autonomy leads people to barely perceive the AI’s usefulness.
• Too much autonomy triggers resistance, because the feeling of one’s own control is lost.
• The highest acceptance and the greatest effects arise with medium autonomy: here, users experience the machine as support, not as a substitute.

A decisive factor in this is the sense of self-efficacy (“self-efficacy”), that is, the conviction that one can influence one’s own decision oneself. The stronger this feeling remains, the more readily people accept algorithmic help. If it is taken away from them, they react with mistrust or rejection.

This insight can be applied well beyond the consumption context: in recruiting, in finance or in project management, the same holds true: the balance between technical assistance and human control determines acceptance and effectiveness.

Who sees which options, how strongly are they prioritised, and who ultimately owns the decision?

Empirical findings

eCommerce: perception of transparency and reliability

In a study by Sourabh Satish Zanwar (2023), the researchers analysed how the transparency and reliability of an algorithm influences users’ trust and decision quality. In several experimental setups, participants received recommendations whose origin was either explained (“transparent condition”) or obscured (“non-transparent condition”).

The result was ambivalent: participants trusted even non-transparent systems as long as the suggestions worked in the short term. In the long run, however, acceptance dropped markedly as soon as a faulty result occurred. Transparent systems, by contrast, were rated more positively even after errors, because the users could understand why the error happened.

This observation aligns with findings from cognitive psychology: trust in technical systems arises less through perfection than through explainability. For companies this means that an explanatory interface or a visible justification (“We recommend X because you used Y”) can stabilise trust, especially when errors are unavoidable.

The influence of popularity and social proof

A further empirical focus lies on the principle of popularity. In Hazrati’s study Choice Models and Recommender Systems Effects on Users’ Choices (2024), large-scale simulations showed that even small cues of popularity, such as “Popular with users” or “Top rating”, have a significant influence on choice distributions.

The intriguing part: the researchers were able to demonstrate that users’ behaviour does not change in proportion to the actual quality of the products, but in proportion to their perceived social relevance. Popularity thus becomes a decision driver in itself, independent of objective usefulness.

This gives rise to a feedback effect: recommender systems promote what is already popular, which makes it even more visible and therefore even more popular. This “feedback loop” can lead, in a short time, to a market concentration in which a few products, providers or pieces of content dominate.

Diversity, quality and long-term effects

While many studies examine short-term decision processes, Hazrati & Ricci (2021) engaged with the area of “Quality and Diversity of Recommender Systems Users’ Choices” and, in doing so, put long-term development under the microscope. In repeated simulations it emerged that, while strongly personalised systems generate higher satisfaction in the short term, in the long term they lead to homogenisation.

Users increasingly received similar recommendations, yet lost access to “edge options”, that is, products, ideas or information outside their previous patterns. This effect is also observable in other areas: music streaming services such as Spotify increasingly show users songs that are stylistically very close to their existing taste. What basically sounds great is in truth counterproductive. Because as a result, the discovery effect declines.

When the competence for discovery is lost, people become less willing to engage with new things on their own, especially when systems mirror their preferences too closely. Companies that deploy recommendation mechanisms in their processes and systems must deliberately build in counterweights: random options, “discover something new” functions or curated diversity suggestions.

Trust, control and emotional resonance

In a study by Glikson & Woolley (2023) on Human-AI Interaction, it emerged that trust in AI systems depends less on technical performance metrics than on the perceived social competence of the machine.

Participants described “likeable” systems (for example, with empathetic language or personalised forms of address) as more credible and more helpful, even when they were objectively less accurate. This paradox points to a deeper insight: people react to machines according to the same psychological patterns as they do to other people. They look for reliability, explainability and emotional resonance.

Implications for entrepreneurs

All the empirical findings can be traced back to one central pattern: it is not the existence of an algorithm that determines its success, but its perceived role in the decision process.

AI systems have long since stopped being neutral tools. They are co-players. They negotiate along with us about what becomes visible, what seems plausible, and what even appears to us as a “reasonable” decision in the first place.

Explainability beats perfection

Research shows that, in the long run, people accept faulty but explainable systems more strongly than flawless but non-transparent ones. This insight is almost paradoxical, and at the same time highly relevant for business. When systems are allowed to make mistakes, but those mistakes remain comprehensible, a new form of functional trust emerges.

Companies should therefore not strive for the perfect machine, but for the explainable machine. “Explainable AI” (XAI) is not an ethical add-on but a business advantage: it reduces support effort, strengthens brand trust and can even become a competitive advantage from a regulatory standpoint (for example, through the EU AI Act).

It is not the cleverest AI that wins, but the one whose errors people are most willing to forgive.

Balancing influence and freedom

As Fan & Liu (2022) showed, algorithmic autonomy works most strongly when it does not patronise but supports. From this, a principle can be derived for entrepreneurs: assist, but do not take over.

In practice, this means thinking of the design of the AI interaction as a cooperative relationship. Examples:

• A recruiting system that makes suggestions but leaves room for decision is more readily accepted than one that selects automatically.
• A price-optimisation AI that justifies its calculation (“market demand + storage costs + seasonality”) is perceived as a partner, not as control.

Too much algorithmic paternalism leads to a form of digital powerlessness: you believe you are saving time, but you lose the inner connection to your own decision.

To automate autonomy is to abolish self-determination.

Diversity is economic resilience

Systems that allow diversity are more stable in the long run. The more homogeneous recommendations become, the less need users have to discover new products or ideas. For companies this means: anyone who bets on algorithmic efficiency without securing diversity is breeding a monoculture. And monocultures die of themselves.

Recommender systems should deliberately build in disruptive factors: random offerings, curated deviations, logics of contradiction. In economics this is called Algorithmic Serendipity, the deliberate planning-in of surprise.

Without chance, innovation withers. The most efficient algorithm is the deadliest enemy of creativity.

Emotional intelligence counts

The study by Glikson & Woolley (2023) makes clear that people fundamentally do not trust machines because they can compute, but because they listen. The “how” of a recommendation often counts for more than the “what”.

An AI that explains, simulates empathy or merely sounds friendly generates emotional resonance. This is not a sentimental luxury: in times of growing AI penetration, emotional perception determines acceptance rates, customer satisfaction and compliance within the company.

Power, responsibility and “behavioural design”

Algorithmic decision architectures are instruments of power. Whoever steers what is seen, recommended or omitted shapes perception. Globally. Recommender systems are nothing other than institutionalised nudges, digital and scalable.

For entrepreneurs, this gives rise to a twofold responsibility: to understand the power of nudging and to use it ethically, and to design these mechanisms deliberately instead of tacitly outsourcing them to AI providers.

The future of the economy does not depend on who advertises the loudest, but on who influences the most quietly.

Limitations, open questions and outlook

Despite the empirical grounding, several limits and open fields of research remain:

• Limited field studies: Most studies stem from controlled laboratory or online experiments. Real corporate environments with multiple decisions, time pressure and political dynamics are more complex. Long-term data is lacking here.
• Cultural differences: Studies show that cultural factors (for example, power distance, individualism) massively influence the acceptance of algorithmic systems. A design that works in Germany can trigger mistrust in Japan.
• Data ethics and regulation: With the EU AI Act, a legal framework is emerging that demands algorithmic transparency. Yet regulatory compliance does not yet guarantee acceptance. What is decisive is the human experience of fairness.
• The future of leadership: When AI systems co-shape strategic decisions, that changes the role of executives. Do they become curators of machine decisions? Or translators between human and system?
• A societal question: When decision systems take decisions off our hands, responsibility shifts. Is it the developer, the entrepreneur, the algorithm or the user who only clicks?

Conclusion

Recommender systems and algorithmic decision autonomy are not peripheral topics of digitalisation; they are its inner mechanism. They define how power, attention and trust are distributed in the digital age.

For entrepreneurs, this yields a simple but uncomfortable truth:

Anyone who does not understand how AI steers decisions will soon no longer understand how their customers decide.

The way forward therefore does not lead through blind automation, but through deliberate co-creation: AI systems that do not replace decisions but extend them; algorithms that promote diversity instead of reducing it; and entrepreneurs who do not outsource their responsibility to machines but reinterpret it anew.