Governing Artificial Intelligence, or Finally Learning to Govern?
Eight lessons from Sara Moroni, three ribbons, some uncomfortable questions, and a suspicion: perhaps AI is not only changing work but also the meaning of managing it.
I just finished following the eight chapters of Sara Moroni's masterclass AI Governance Gap, a good forty minutes dedicated to a question that, paradoxically, is still discussed too little, considering the enthusiasm with which we are integrating artificial intelligence into every business process, every application, and possibly even into some activities that worked perfectly well without it. I expected a path on governance, and I found it, but above all, I found myself reflecting on something I know quite well, having spent a significant part of my professional life managing teams, coordinating projects, building software, and, lately, doing more or less the same things with collaborators who do not take vacations, do not ask for raises, and occasionally come up with solutions that no one had asked them for.
The question that Sara starts with is essentially this: how much do we really know about the artificial intelligence being used in our companies? Not how much we use it, not how many licenses we have purchased, not how many employees have discovered that ChatGPT can save them two hours of work, but how effectively we are able to understand its applications, measure its results, recognize its risks, and intervene when something goes wrong. It is a distinction less trivial than it may seem, because in the history of corporate computing, we have often confused the presence of a technology with the ability to use it, and the ability to use it with the ability to govern it. Now that tools arrive even before someone has decided to adopt them, the problem has become particularly evident.
The masterclass is available for free at masterclassaigov.it and I believe it deserves to be followed, especially by those with decision-making responsibilities in an organization. Not because it delivers a magic formula, which fortunately Sara does not pretend to do, but because it raises a series of questions to which it would be appropriate to know the answers without having to urgently convene a meeting.
Three ribbons for a running problem
The heart of the reasoning is represented by the three ACG ribbons: Adoption, Compliance, and Governance. Adoption describes how much AI has penetrated the organization, compliance concerns adherence to applicable obligations, and governance ties together responsibilities, processes, results, and adaptability. Sara insists on an important distinction: governance is not simply the third office to add to the other two, but the system that allows the organization not to lose track while everything else continues to move.
It seems to me an effective representation, especially for those companies that confuse subscribing to some service and drafting a policy with a digital transformation strategy. The case of the SaaS company presented in the fourth lesson makes the concept particularly understandable: a technologically competent organization, in which artificial intelligence produces genuine results, may discover that it does not possess a shared vision of its uses. The competence of individuals does not automatically transform into collective capability. It is a problem that anyone who has managed groups of developers should immediately recognize, because you can have a team made up of excellent people and a project that still proceeds without a sufficiently clear direction.
The other case, that of the MedTech company, highlights the opposite problem. A company can be extremely rigorous in governing the artificial intelligence embedded in its products and much less attentive to the tools used daily in the offices. The culture of control exists, but it has remained confined within a perimeter. It is an interesting observation because it demonstrates how dangerous it is to talk about corporate maturity as if it were a uniform property, a grade to assign to the entire organization, while in reality, the same company can be extraordinarily disciplined in one department and surprisingly improvisational in the one next door.
So far, I fundamentally agree. Where I begin to question is the speed of the three ribbons, because the representation works very well to show that they must proceed together, but the issue that seems most urgent to me today is to understand what happens when one of them accelerates to the point of making it almost impossible for the others to keep up.
I have already seen this problem, only it used to have a desk
In recent months, I have been using AI agents to simultaneously develop different functionalities and products. I do not limit myself to asking for code and pasting it somewhere hoping it works: I set objectives, define constraints, verify results, intervene on architectural decisions, correct errors, and sometimes discover that one of the agents has found a better solution than mine. I have realized that this activity resembles less and less traditional programming and more and more what I used to do when I managed groups of developers. The fundamental difference is that a work sequence that previously required days can now be completed in the time it takes to make a coffee, and if in the meantime I am following another project, I might find myself with several technical decisions already transformed into concrete results.
This does not mean that the agents are employees, nor that they possess autonomy in the human sense of the term. However, it does mean that the organizational problem of delegation has separated from the human speed with which we were used to managing it. And here arises a question I would like to add to Sara's path: are we building a governance capable of accompanying this acceleration, or are we trying to impose the same decision-making times of traditional organizations on radically new systems?
Sara, moreover, does not propose a static system at all. In the fifth chapter, she presents an eight-step framework, from planning to updating, passing through resources, KPIs, implementation, training, initiation, and monitoring. She talks about continuous controls, periodic reviews, and even the need to recognize unexpected benefits. It is an important point because it would not make sense to blame her for a static conception that she herself is explicitly trying to overcome.
I would take a further step, distinguishing two speeds of governance. The first is organizational, in which objectives, responsibilities, budgets, limits, and risk acceptance criteria are established. The second is operational, in which a system must adhere to those limits while working, without waiting for new authorization for each single action. An agent that can read a repository, propose changes, and run tests in an isolated environment does not need the same level of preventive supervision as an agent authorized to publish code in production, send data to third parties, or make decisions that affect people. The distinction should not simply depend on the fact that both are called AI agents, but on the concrete power we have entrusted to them.
In other words, we don’t necessarily have to monitor every movement. We need to know what movement space we have authorized, what consequences we can sustain, and who can intervene when that perimeter is exceeded. Which, thinking about it, is a fairly reasonable definition of management even when the collaborators are human beings.
Responsibility is not a signature on the form
One of the points I consider most important in the masterclass concerns the attribution of responsibilities. Sara rightly observes that we cannot wait for a mistake to ask who is responsible for it. I agree, but I believe that with autonomous agents the issue deserves further exploration, because in organizations it is relatively easy to assign formal responsibility and much more difficult to give the person assuming it the actual power to exercise it.
Imagine a manager who is entrusted with the supervision of an AI system that interacts with different tools, modifies information, consults archives, and produces operational decisions. We can write their name in a document, include them in an organizational chart, and declare that they remain responsible for the functioning of the system. But can they interrupt it? Do they know which tools have been used? Can they reconstruct the sequence that produced an unexpected result? Do they know the limits of the permissions granted? If they discover a problem, can they revoke those accesses without waiting for the intervention of five other people?
If the answer is no, we may have identified the recipient of the consequences, but we have not necessarily built an exercisable responsibility. And this is one of the points where, in my opinion, modern governance should be particularly concrete: less emphasis on the signature, more attention to the ability to observe, understand, and intervene.
Of course, not all operations are reversible. Confidential information sent outside does not become secret again just because we pressed a button, just as a decision that has affected someone’s rights does not become harmless just because the system keeps a record of activities. Delegation therefore requires boundaries proportional to the nature of the consequences, not an indiscriminate trust in the possibility of remedying later.
Governance also has a cost. And it should account for its work
Another aspect I appreciated in Sara's journey is the attention to the overall cost of adoption. Not just licenses and infrastructure, but training, learning, skill maintenance, and the time needed to understand and use the tools. It’s a perspective I share, and for this reason, I would ask governance the same question we ask artificial intelligence: how much does it cost, what does it produce, and how do we know if it is working?
Because even a rule consumes resources. It requires someone to define it, someone to interpret it, someone to verify compliance, someone to update the documentation, and frequently, someone else to wait for the result of all these activities before being able to start working. A rule can be indispensable and costly at the same time, of course, but for this reason, there should be a way to evaluate its effectiveness, rather than considering its mere presence as a demonstration of good management.
In the framework, Sara rightly rejects the idea of universal KPIs, because different companies pursue different goals. I share the caution, but I would accompany it with some very concrete examples of measurement. Not just how many use AI, but how much net value is produced after considering checks, errors, and rework. Not just how many procedures have been defined, but how many problematic situations we are actually capable of recognizing and containing. And not just how many risks we have avoided, but also how long it takes us to authorize an experiment, how much our controls cost, and how many initiatives are abandoned because the planned path is disproportionate to their scope.
This last point requires caution: not every prevented experiment represents a lost opportunity. Sometimes it simply means having avoided a bad idea. But the cost of governance should become as visible as that of the governed technology, otherwise, we risk building organizations capable of perfectly documenting everything they can no longer do.
And this is where one of the most delicate issues arises: Shadow AI. Sara deserves credit for reminding us that those who introduce a tool without authorization do not necessarily do so out of irresponsibility, but often because they are trying to work better. It is a fundamental distinction. If an organization makes the official path impractical, it should not be surprised when someone seeks an alternative. The answer cannot be just to add prohibitions, but to understand why that need cannot find a governed solution. Good governance should reduce the need to hide, not just improve the ability to discover who is hiding.
From the framework to Monday morning
If I had to suggest an evolution of the masterclass, I would not add any more large theoretical categories. The path is already clear enough, and the eight steps constitute a comprehensible structure. I would rather like to see one of the two business cases taken beyond diagnosis, to operational decisions and results. In the MedTech case, for example, it would be interesting to follow what happens in the weeks following the discovery of the problem: which tools are first cataloged, which accesses are limited, which uses can continue, who makes the decisions, which metrics are introduced, and what results emerge after a month.
It wouldn’t require another treatise, but a complete example, possibly with some numbers and some difficult decisions. Because the transition from awareness to action is exactly the point where many good organizational models encounter the resistance of reality. Reality, notoriously, has not read the framework, does not participate in alignment meetings, and tends to present problems that are little respectful of the boxes we have prepared to classify them.
I would also add a form of self-diagnosis, a few questions that an entrepreneur could answer without resorting to a consultant: what AI systems are influencing my organization’s decisions? What data can they access? Who can modify their permissions? What would happen if tomorrow one of these systems produced a significant error? And above all, how long would it take us to notice?
These are questions that would also have the merit of distinguishing the governance told from the governance practiced.
Perhaps we need to learn to govern differently
At the end of the eight videos, what impressed me the most is not so much the need to introduce governance of artificial intelligence, which would be hard to disagree with, but the possibility that this technology might finally force us to discuss how we govern organizations in general.
For decades, we have built structures in which the ability to control was often confused with the number of authorized steps, responsibility with position in the organizational chart, and coordination with the need to continuously know what everyone was doing. Those who have managed people know the fatigue of these mechanisms and also know their paradox: when they work too well, they slow down work; when they work too little, they stop providing direction. The true craft of management lies in finding a balance that is not only valid on paper.
With AI agents, that balance becomes even more important, but also more interesting. We can design systems in which many controls are embedded in permissions, architectures, operational thresholds, and the ability to halt a process, leaving people the task of setting goals, evaluating consequences, and making the decisions that truly matter. It is not the abolition of human responsibility, but a more precise placement of it. It does not mean giving up control; it means stopping the identification of control necessarily with continuous authorization.
And perhaps we should apply the same logic to governance rules: periodically checking not only that they are being followed but also that they still make sense, that they produce the expected effects, and that their cost remains proportionate to the risks they intend to contain. In this way, the governance of the organization would become a process capable of learning, instead of a structure that demands to teach others without ever undergoing an examination.
This, in my opinion, is the most interesting reflection that Sara Moroni's masterclass allows us to initiate. Her three tapes help recognize a real problem, and the framework offers a path to address it. The next step, on which I would like to continue reflecting, concerns the ability to distribute autonomy without losing responsibility, to make control effective without making it omnipresent, to build organizations capable of changing without having to ask for permission from themselves every time.
Because if artificial intelligence can change the way we work, it would be rather curious to use it to make the same organizational mechanisms that have prevented people from working well for years even more efficient.
And perhaps successful governance should not only be measured by how many mistakes it can prevent but also by how many good things it finally allows to happen.
The masterclass
Those who wish to delve deeper can follow for free the eight chapters of AI Governance Gap, Sara Moroni's masterclass dedicated to the relationship between the adoption of artificial intelligence, regulatory compliance, and the governance of organizations.
AI Governance Gap: access the masterclass
About 45 minutes, two business cases, and a model in eight steps. A good starting point for those who want to ask better questions before purchasing yet another AI license.
Is your project stuck at a technical hurdle?
Book 15 minutes