Interview with Massimiliano Ferraris
As AI moves upstream into how decisions are formed, are organizations still truly in control or simply operating within systems they no longer fully understand?
Governance expert Massimiliano Ferraris reflects on how artificial intelligence is reshaping decision-making, raising urgent questions about sovereignty, responsibility, and the quality of organizational choices.
What is drawing your attention to how organizations are being governed in the age of AI right now, and what feels most urgent about it today?
What defines this moment is not the acceleration of artificial intelligence, but the widening gap between what organizations are technically capable of doing and what they are structurally capable of governing.
Across industries, one observes a recurring pattern. Artificial intelligence is deployed, adoption is rapid, and measurable efficiency gains follow. Processes become faster, outputs become more consistent, and internal coordination becomes more fluid. By every conventional metric, these initiatives are considered successful. And yet, when one asks a more fundamental question, whether the quality of decisions has actually improved, the answer is often uncertain, and in many cases, unknowable.
This is not a transitional misalignment. It reflects a deeper structural condition. Modern organizations are optimized to measure what is immediately visible: productivity, speed, adoption, and cost reduction. Decision quality, by contrast, is delayed in its manifestation, difficult to isolate, and often only observable in moments of stress or discontinuity. Because it resists simplification, it is rarely measured. Because it is not measured, it is not governed.
The urgency lies in recognizing that artificial intelligence is not entering organizations as a tool. It is reconfiguring the architecture within which decisions are produced.
The consequence is a paradox that increasingly characterizes advanced organizations. Systems become more efficient and more data-rich, while the underlying capacity to decide does not evolve at the same pace. In certain configurations, it deteriorates. What appears as technological progress can conceal a progressive erosion of the quality of governance.
What makes this condition particularly dangerous is that it is self-reinforcing. Organizations that evaluate success through adoption metrics conclude that they are advancing, even as the structure of their decision-making remains unchanged. In doing so, they fail to capture the full value of artificial intelligence. They construct, incrementally, the conditions for their own relative decline.
The urgency lies in recognizing that artificial intelligence is not entering organizations as a tool. It is reconfiguring the architecture within which decisions are produced. It is shifting the locus of control upstream, into the design of the decision environment itself. Governance, however, remains anchored downstream, focused on outputs, compliance, and ex-post validation. The misalignment between these two layers is no longer marginal. It is structural, and it is widening.
How has the way decisions are made inside organizations changed with AI becoming more present in everyday work?
The transformation is frequently described as an increase in efficiency. This description is insufficient. The deeper shift concerns the nature of decision-making itself.
Historically, decisions were constructed. Human actors defined problems, generated alternatives, and navigated trade-offs through deliberation. Artificial intelligence introduces a different sequence. Systems generate options, structure narratives, filter alternatives, and present coherent outputs before human reasoning begins.
The effect is subtle but profound. Decisions are no longer constructed through a process of reasoning. They are encountered as outputs within a pre-configured space of possibilities.
This shift alters not only the speed of decision-making, but its location. The critical moment no longer resides in the act of choosing among visible options. It resides in the prior configuration of the option space itself. By the time a decision reaches an executive or a board, it has already been shaped by a system that has defined what is relevant, what is comparable, and what is worth considering.
What is often overlooked is that this transformation does not simply affect how decisions are made. It affects what can be decided at all. When the system filters the option space, it is not accelerating deliberation. It is redefining the cognitive perimeter within which deliberation occurs. The most consequential effect is therefore not the recommendation it produces, but the set of alternatives it renders invisible.
In this sense, decision-making is no longer a process of construction. It becomes a process of interaction with a space that has already been constructed elsewhere.
How would you explain decision sovereignty in simple terms, and why do you think it matters now more than ever?
Decision sovereignty is commonly understood as the authority to approve or reject a decision. In the context of artificial intelligence, this definition is no longer sufficient.
To be sovereign is not simply to choose among available options. It is to define the conditions under which those options emerge. It is to determine the structure of the decision environment itself.
Artificial intelligence intervenes precisely at this level. It does not merely support decisions. It configures the set of possibilities that appear viable, reasonable, or even imaginable. It operates at the level of what can be thought, not merely what can be chosen.
This is what can be described as the governance of the architecture of the thinkable.
Artificial intelligence intervenes precisely at this level. It does not merely support decisions. It configures the set of possibilities that appear viable, reasonable, or even imaginable.
The risk that follows is not primarily that of incorrect decisions. It is the progressive contraction of the decision space. Certain alternatives become less visible, less accessible, or entirely absent, not through explicit prohibition, but through the dynamics of optimization embedded in the system.
In such a context, sovereignty can be lost without any formal transfer of authority. Decisions continue to be approved by human actors. Responsibility remains formally assigned. But the space within which those decisions are made has already been shaped elsewhere.
The most critical governance failure of this era is not that organizations make poor choices. It is that they increasingly operate within decision environments they did not consciously design. And sovereignty, once displaced at that level, cannot be recovered at the point of approval.
In your view, where should human judgment always remain central, even as AI becomes more involved in decision-making?
The relevant question is not which decisions should remain human. It is at which level human judgment must remain operative.
Human judgment must retain primacy where the structure of the decision environment is defined. This includes the formulation of the problem, the selection of objective functions, and the capacity to reconstruct the logic of decisions.
The definition of a problem is not a technical act. It is an act of institutional meaning-making. It determines what is relevant, what is excluded, and what constitutes success. Similarly, the selection of an objective function is not neutral. It encodes priorities and trade-offs into the system, often in ways that are not immediately visible in the output.
Perhaps most critically, human judgment must retain the ability to reconstruct decisions. This includes understanding not only why a particular option was selected, but why other options were excluded. Without this capacity, responsibility becomes purely formal. It is attached to outcomes that cannot be fully understood.
The bigger risk is not automation. It is the delegation of cognitive framing. When systems define not only the answer but the question, human judgment becomes reactive. It operates within a space that has already been configured. At that point, organizations do not lose control over execution. They lose control over meaning.
Where do organizations typically start to struggle when they begin using AI to support or influence decisions?
Organizations struggle because they interpret artificial intelligence as an instrument of efficiency rather than as a transformation of cognitive infrastructure.
This misinterpretation interacts with a pre-existing condition. Most organizations operate within an equilibrium in which execution is rewarded, predictability is valued, and the interrogation of underlying assumptions carries an implicit cost. This equilibrium is not accidental. It is the rational outcome of incentive structures.
In such an environment, the cognitive investment required to understand and govern AI systems, to interrogate their assumptions, to redesign processes accordingly, does not produce immediate, measurable returns. It is therefore rationally avoided.
Artificial intelligence does not disrupt this equilibrium. It amplifies it. It increases the speed of output, the coherence of recommendations, and the perceived reliability of systems. At the same time, it reduces the perceived necessity of interrogation.
The result is a form of scaled stability. Organizations become more efficient at doing what they were already doing. They do not become more capable of doing what they ought to do.
What parts of traditional ways of running organizations are no longer working well in today’s environment?
The elements that are failing share a common assumption: that decision-making is observable, sequential, and human-originated.
Traditional governance frameworks focus on outputs. They evaluate whether decisions are correct, compliant, or aligned with policy. This presupposes that the decision itself is the relevant unit of analysis. In AI-mediated environments, this assumption no longer holds. The relevant unit is the environment that produces the decision.
This misalignment is compounded by a deeper structural condition that can be described as cognitive governance asymmetry.
The cost of configuring a decision environment through algorithmic systems is systematically lower, and decreasing over time, than the cost of governing that configuration through institutional processes. This is not a temporary gap. It is a structural property of the relationship between computation and deliberation.
Automation operates at computational speed. Governance operates at cognitive and social speed. The divergence between these two dynamics produces a predictable outcome. Control tends to migrate toward those who configure the system rather than those who supervise it.
This is not a failure of governance. It is the consequence of governing at the wrong level.
When decisions are shaped by both people and AI, how should responsibility be clearly understood and assigned?
Responsibility, in its current form, is anchored to the moment of approval. It assumes that the individual who approves a decision possesses sufficient visibility and understanding to be meaningfully accountable.
In AI-mediated environments, this assumption becomes increasingly fragile. Decisions are generated within systems that filter options, embed trade-offs, and optimize according to functions that may not be fully transparent. The individual who approves the decision may not have access to the full structure that produced it.
This creates a divergence between formal responsibility and effective control.
Responsibility does not disappear. It is preserved in form. But it risks being emptied of substance. The signature remains. The understanding behind it does not.
To restore coherence, responsibility must extend upstream. It must encompass the governance of the decision environment itself. This requires the capacity to interrogate what was not considered, to understand the objectives that guide system behavior, and to define the domains in which the configuration of the decision space must remain under human control.
Absent this shift, accountability becomes symbolic. It exists formally, but no longer corresponds to actual control.
How do you think decision-making inside organizations will continue to change as AI becomes even more deeply embedded in everyday work?
The evolution is not linear. It is bifurcating.
Some organizations will continue to treat artificial intelligence as a productivity layer. They will achieve efficiency gains, optimize processes, and improve execution. Their structures will remain largely unchanged. These organizations will appear successful, but they will accumulate a latent vulnerability: the inability to adapt their decision architecture as conditions evolve.
Organizations will continue to treat artificial intelligence as a productivity layer. They will achieve efficiency gains, optimize processes, and improve execution. Their structures will remain largely unchanged.
Other organizations will recognize that artificial intelligence operates at a structural level. They will redesign their processes, integrate AI into the generation of decisions, and develop governance mechanisms capable of operating upstream. These organizations will not simply be more efficient. They will develop a distinct form of capacity: the ability to shape and control their own decision environments.
Over time, this divergence will not manifest merely as a difference in performance. It will manifest as a difference in cognitive capacity. Some organizations will retain the ability to redefine the conditions under which decisions are made. Others will remain confined within decision environments they no longer fully understand.
Closing Reflection
The central question facing organizations today is not how to use artificial intelligence.
It is whether they retain control over the conditions under which their decisions become possible.
The question is not whether organizations will continue to make decisions. They will.
The question is whether those decisions will still be meaningfully attributable to them.
As artificial intelligence systems move upstream, into the configuration of decision environments, the locus of control shifts in ways that are not formally recognized. Authority remains where it has always been. But the structure within which that authority operates is no longer fully visible.
In that condition, governance does not collapse. It persists, but at a different level of effectiveness.
And over time, the distinction between deciding and inheriting decisions becomes increasingly difficult to observe from within the system itself.
