A marketing team has to decide how to position a product. There are two options on the table. The CEO uploads the documents and asks AI for an analysis. The model produces a structured report: coherent scenarios and plausible recommendations steer the decision on the basis of that analysis.
The crucial point is that power lies not in the final decision, but in defining the architecture of the decision. Whoever establishes which variables are relevant, which metrics carry more weight and which risk thresholds are acceptable has already steered the outcome. Choosing between options becomes secondary to the preliminary selection of the field of possibilities. AI defines which alternatives deserve to be taken seriously. It is in this pre-decision framing that the organisation’s cognitive sovereignty shifts.
The same dynamic recurs when a system identifies the ideal candidate for a job in a company: the boundaries have already been drawn. To depart from them, HR would have to explicitly assume the risk of contradicting an analysis perceived as more “objective”. In this case, AI legitimises a direction in advance, conditioning the choice.
The phenomenon cuts across business functions. In legal matters, for example, an action that the model has classified as difficult to sustain is unlikely to be proposed; likewise, options that fall outside the initial analysis are rarely explored.
In each case, the system does not decide on people’s behalf: it sets the boundaries of what seems reasonable to decide. This means that AI, too, was present at those meetings, alongside everyone else, recognised and accepted as a legitimate interlocutor. Not because it is formally credited with a mind, but because its contribution is taken as legitimate and authoritative.
Acceptable responsibility is shifting
Organisations that care about their future must begin to confront a potentially destabilising transformation: the point at which it is psychologically acceptable for organisational responsibility to stop is shifting towards AI.
Indeed, increasingly often, tracing responsibility for a decision leads to the conclusion that “the model suggested it”, that “given the data, it was inevitable” and that “the system’s guidance left no room”.
What is striking is that these empty phrases produce a sense of logical closure and cause no unease. On the contrary, they leave people relieved of the burden of deciding, as though it really were enough to have established that responsibility should be attributed to this sort of technological subject that is AI.
From shame to cognitive dependence
Until recently, writing a text with ChatGPT was something to feel a little ashamed of. Now the situation has changed. Many managers consult AI to feel fully informed. Not doing so feels like a kind of negligence. Relying solely on one’s own judgement and that of one’s peers can even seem superficial.
“If I have a complex organisational problem and do not consult the best-informed system available, perhaps I am not really doing the best I can for the company.”
After all, whenever we have been unable to resolve an issue, we have always moved up a level: we consult someone who knows more. Today, for many people, that “someone who knows more” is no longer the boss or a senior colleague. It is the general-purpose model, which thereby becomes the subject of legitimation: it does not carry the same emotional burden as talking to the boss, nor does it defend personal positions and interests as certain colleagues do.
This means that the decisive transition has taken place: from the days when we debated whether AI was an acceptable tool to the present, when, from an accepted tool, it has become a super-competent colleague occupying the functional position of the organisation’s cognitive authority.
The automatic application of legitimation
Anyone who has interacted at length with an AI model knows how skilfully the system argues, anticipates objections or reformulates. The fact is that this level of linguistic ability activates an automatic cognitive mechanism that Daniel Dennett described as the intentional stance: we attribute intentionality and a mind to any entity that displays coherent, goal-directed behaviour.
In any case, the only way we have ever had of gaining access to other people’s minds has been to infer them from their speech acts. In the context in which our cognitive system developed, linguistic production was always a reliable indicator of the presence of a mind. Today, that indicator no longer distinguishes between them. Language models replicate precisely those signals that, through cognitive habit, we associate with a subject.
As Nicholas Epley has shown, anthropomorphism is a systematic process: when an entity communicates in a responsive, linguistic and apparently goal-directed way, we tend to attribute a mind and intentions to it, for reasons of cognitive economy and efficiency.
The point, however, is not merely that we thereby attribute a mind, a dynamic extensively analysed in the philosophy of mind, but that attributing a mind produces legitimation. And legitimation is the basis of organisational responsibility.
If an interlocutor appears competent and coherent, stopping at their conclusions becomes psychologically and socially acceptable. When AI intervenes first, it defines the framing: in effect, it establishes which variables count and which do not. What falls outside those boundaries struggles to emerge; it requires an additional effort and often never surfaces at all. The decision that follows is therefore coherent within that frame, but not necessarily the best one within other possible frames.
Where responsibility stops
In many organisations, the decision-making sequence has already shifted: first, people consult the model to get an idea, then they discuss it in a meeting, each with their own generated model, and then they converge on the most convincing proposal. In this arrangement, AI defines what is relevant before human discussion even begins. Criteria, variables and thresholds have already been selected. The discussion takes place within boundaries that have not been deliberated, and leadership risks being reduced to the coherent execution of premises it has neither conceived nor defined.
When this sequence becomes the norm, it produces three structural effects. The first is premature convergence: participants arrive already aligned within a narrow range of options, and dissent diminishes not through genuine agreement, but because the possibilities are limited. The second is cognitive homogenisation: different organisations end up reasoning within similar frameworks, because general-purpose models propose recurring analytical structures and priorities. The third is the transformation of leadership into a function of logical compliance: the leader no longer defines the criteria, but consistently applies a structure already given.
All this can be avoided by reversing the sequence. Before turning to models, the group must make its decision-making criteria explicit: which variables count, which values are non-negotiable, which risk thresholds are acceptable and which relevant dimensions we need to explore that AI cannot see.
Only once these boundaries have been drawn can the system be used to explore scenarios and alternatives. At this point, the decision can return to people, who finally take it within consciously established boundaries.
If the boundaries of decision-making are outsourced, responsibility tends to follow. Not because anyone wishes to evade it, but because what appears rationally inevitable ceases to be questioned. Organisations learn not only from the decisions they make, but from the criteria they discuss. When criteria are no longer subject to explicit deliberation, the culture narrows and the governing function weakens.
This article was first published on EconomyMagazine.it at this address: https://www.economymagazine.it/il-collega-super-competente/