The AI Act regulates high-risk systems, and Gartner promises an explosion in the number of AI agents by 2028. Governing them becomes an extensive managerial skill, not just for the CIO.
Two marks are enough to measure the tilt in progress. On August 2, 2026, the European AI Act enters its most restrictive phase and imposes on high-risk systems documented human supervision and traceability of their decisions, a framework designed for machines that act, and no longer just respond. At the same time, a Gartner study anticipates that large American companies will each deploy more than 150,000 AI agents by 2028, compared to around fifteen on average last year. This acceleration is not just a matter of volume: it shifts a problem that we thought was technical towards a problem of governance, that of knowing who decides what an agent is authorized to do and who is responsible for their errors. This question can no longer remain the exclusive concern of information systems departments.
An autonomy that already escapes individual control
An AI agent plans, decides and acts, alone or in interaction with other agents. This growing autonomy produces effects that no one explicitly anticipated, and many pilot projects are already abandoned due to lack of demonstrated value or sufficient control. Agents responsible for optimizing advertising auctions can, by adjusting each other, saturate the same channels and drive up costs. Pricing agents can converge on identical prices through pure algorithmic mimicry. In HR functions, recommendation agents can develop collective biases and rule out certain profiles, without any isolated error being identifiable. Validating an agent before its deployment in no way guarantees its behavior once in production, where deviations, inefficient loops or overconsumption of resources often appear over time. Managing this type of system requires governing interactions and monitoring overall dynamics, a skill that neither software engineering nor compliance auditing, taken in isolation, covers.
A responsibility that goes beyond IT management alone
Some companies saw their budget dedicated to generative AI completely consumed in a few months, before introducing spending caps per user. The example illustrates a broader phenomenon: the governance of agents becomes a subject of financial as well as technical management. But the issue goes beyond the cost. An agent who automates part of the calculation of insurance pricing, or who sorts applications in a recruitment process, incurs legal liability and the reputation of the organization well beyond the usual scope of the IT department. Deciding which agents we authorize, with what level of autonomy, on which critical processes and according to which chain of responsibility in the event of an error: these are decisions made by general management, business departments and human resources as much as IT.
A major challenge in training
This same question should already occupy higher education establishments. Training future managers or engineers in the use of generative AI tools is becoming a standard with numerous initiatives over the past 2 years; preparing them to govern systems that decide without constant supervision is much less so. This requires teaching reasoned delegation: how far to entrust a decision to an agent, when to take control, how to read the indicators that signal a deviation, and how to articulate technical autonomy and managerial responsibility. It is no longer a digital skill among others: it is a governance skill in its own right, which every future leader, manager or engineer must master, whatever their original profession.
The governance of AI agents is not another technical project to be entrusted to CIOs. It is a managerial skill, which involves the strategy, compliance and reputation of the entire organization. The companies that will gain the advantage will not be those that have deployed the most agents, but those that will be able to decide, at any moment, how far to delegate the decision to them. This requirement goes beyond just the business world: it questions our collective capacity to remain masters of the systems we set in motion. Governing autonomy will be, in the coming decade, a skill as decisive as controlling the budget or controlling teams was yesterday.
