In a financial advisory meeting, taking notes, organizing decisions and preparing the follow-up message can consume a significant share of the work. Morgan Stanley has assigned these tasks to a generative system that, with the client’s consent, produces summaries, action items and a draft to be reviewed before it is sent. One advisor estimated a saving of about half an hour per meeting. The half-hour matters, of course, but it does not exhaust the change: more attention remains available during the conversation and, after the meeting, less energy is absorbed by the administrative upkeep of the relationship. Along with productive capacity, the technology thus gives back a portion of mental presence that management must learn to manage.
Something similar, on a different organizational level, has occurred at IKEA. Between 2021 and 2023, the Billie chatbot handled 3.2 million interactions, approximately 47% of all requests received, while 8,500 call-centre employees were retrained for remote interior design, digital sales and the management of complex cases. Part of the work was transferred to the machine. People were redirected towards activities in which experience, listening and the interpretation of needs that have not been fully articulated continue to carry tangible weight.
For businesses, the problem begins after the saving has been achieved. Every technology promises efficiency; artificial intelligence, however, makes available something more difficult to account for: time and attention that can be devoted to new products, training and the oversight of automated decisions, or rapidly reabsorbed by other procedures, messages and meetings. The hour that is returned has no predetermined economic value. It acquires one according to the decision that determines how it is used, and that decision rarely appears in traditional measurement systems. Yet it affects the ability to innovate, retain expertise and build reliable relationships with customers and suppliers.
The speed of adoption reveals what is at stake. In Italy, among companies with at least ten employees, the use of AI rose from 8.2% in 2024 to 16.4% in 2025; among large companies, the figure reached 53.1%, while among small and medium-sized enterprises it remained at 15.7%. The gap cannot be interpreted solely as a difference in investment capacity. Large organizations are more likely to have well-organized data, formally assigned responsibilities and the expertise required to integrate tools into their processes. Where these conditions are absent, even advanced technology may remain an accessory purchased without a sufficiently well-defined managerial purpose.
The less obvious danger is that artificial intelligence may accelerate the mediocrity already present. A company may produce more documents, respond more quickly and process a greater number of cases without gaining a better understanding of the market. If performance indicators primarily reward volume, part of the bureaucracy becomes faster without the organization becoming genuinely more intelligent. One step is eliminated and others are added; the number of alternatives increases, but this does not mean that they are assessed more carefully. Even a schedule freed from previous tasks may be immediately filled again, until productivity comes to coincide with nothing more than an increase in operational density.
The hours saved therefore represent only a partial measure. It is also necessary to consider the quality of decisions, the errors detected, the knowledge that remains available within the company and working conditions after automation. Not all these elements can easily be incorporated into an indicator, and management does not always possess the tools required to monitor them. Part of AI governance will take shape around this difficulty.
Human capital can be strengthened or eroded by the same technology. In more structured roles, AI spreads effective practices and shortens the learning process for less experienced workers, but the nature of the advantage changes when automatic suggestions begin to replace the exercise of judgement. A form of cognitive dependence then emerges that does not coincide with the system’s occasional errors: it concerns, rather, the habit of no longer reconstructing the path that leads to an answer.
Training will also have to encompass this dimension. Recognizing a plausible but incorrect conclusion, tracing data back to its source, challenging a forecast and understanding when a decision must once again be placed under identifiable human responsibility are not ancillary skills. Turning them into a new catalogue of procedures, however, would not be sufficient: judgement is also formed through experience of cases that do not conform to the expected pattern.
What appears immaterial within the office continues to depend on a very concrete structure. In 2025, private investment in AI in the United States was more than twenty-three times that of China, while the performance gap between the leading models of the two countries had almost closed. Almost all frontier chips still depend on a single major Taiwanese foundry. Data centres consumed approximately 415 terawatt-hours in 2024 and could reach 945 by 2030.
Behind an answer that appears on a screen within a few seconds are energy, semiconductors, cloud infrastructure, cables and jurisdictions. Adopting a model therefore also entails choosing among industrial dependencies, access conditions, security standards and continuity guarantees. For many companies, these aspects remain distant from everyday management, at least until a contractual change, a service interruption or a new regulatory constraint suddenly makes them visible.
Corporate sovereignty does not require every infrastructure to be built within the organization. Rather, it requires identifying what cannot be relinquished without losing autonomy: strategic data, evaluation criteria, organizational memory and the power to halt an automated process now belong to the same governance architecture. With European enforcement and new transparency rules for certain systems and content taking effect on 2 August 2026, traceability is also assuming growing importance for trust and reputation. Compliance therefore enters into the design of the chain of responsibility, rather than remaining a control applied only at the end.
The number of artificial agents introduced will, by itself, reveal little about how advanced a company is. More indicative will be its ability to assign a verifiable purpose to the time made available: part of it may become productivity, while another part may become learning, relationships or oversight. In practice, these purposes will tend to overlap and will not always be easily distinguishable.
The rest will depend on managerial culture, because an algorithm can optimize a task without determining what genuinely deserves an organization’s attention. When execution becomes cheaper and more abundant, judgement, trust and responsibility assume a different weight. Financial statements struggle to represent them, but a growing share of economic value will also depend on how these resources are cultivated and used.