Artificial intelligence in business

Work and agent-based AI: when algorithms enter the organisation

Agent-based AI is not merely automation, but a profound change in processes and leadership, bringing new challenges regarding accountability, collaboration and the promotion of professional expertise

by Gianni Rusconi

28 September 2026

Aggiungi Il Sole 24 Ore
ai preferiti su Google

6' min read

Translated by AI

Versione italiana

6' min read

Translated by AI

Versione italiana

artificial intelligence is rapidly changing the way we work, and according to the prevailing view among various experts, the most significant shift may not be the one relating to individual productivity. With generative AI, we have learnt to interact with systems capable of producing text, images and summaries; now, with the agent-based AI, the relationship is changing once again. The model is no longer simply asked to respond, but to pursue a goal, plan a sequence of actions, utilise data and applications, and act with a certain degree of autonomy. This marks the transition from a support tool to a fully-fledged organisational entity, set to influence decision-making processes, the distribution of responsibilities and the way in which businesses organise their work. An analysis by the Future of Workers Observatory of the Giacomo Brodolini Foundation, based on an international sample of 474 managers, consultants and highly specialised technical professionals, has shed light on this issue.

From technology to processes

According to Paolo Gubitta, full professor of Business Organisation at the University of Padua and co-author of the research, the starting point is precisely this shift in perspective. ‘The new phase,’ the professor explains to *Il Sole 24 Ore*, ‘begins when the AI agent no longer simply works alongside a person but becomes a permanent part of a business process, taking charge of tasks, using data and applications, producing results that serve as input for other activities, and so on.’ Its impact on the organisation, in other words, is not simply a matter of introducing new software, because if the agent becomes part of the process, it is the process itself that needs to be rethought. “It makes sense to start,” Gubitta emphasises once again, “with organisational structures and processes, and therefore to understand how they should function, eliminate redundant steps, clarify responsibilities and decision-making points, and only then decide where and how to integrate the agents.”

This clarification is anything but theoretical, given that agent-based automation requires reliable data, clearly defined access rights and permissions, traceability of activities, quality control of results, and attention to privacy and security. And, no less importantly, it demands closer collaboration between those familiar with the process, the IT functions managing the company’s digital infrastructure, and dedicated control bodies. The most prudent approach, the lecturer suggests, is to start with limited areas and measure the effects on time, costs, quality, errors and service levels.

The manager becomes an ‘orchestrator’

Perhaps the most important consideration, amongst many, concerns the work of those who coordinate people and processes. Whilst the agent is able to analyse information, compare alternatives and formulate proposals, the manager will not necessarily have fewer decisions to make. Rather, they will have more alternatives to choose from. ‘The most interesting change for these roles,’ Gubitta points out in this regard, ‘relates to the exercise of judgement, to that quid that the agent does not possess, at least for the time being. On the one hand, the cost of generating alternative scenarios is reduced; on the other, the cost – including in terms of stress – of choosing which decisions to devote time and judgement to is increasing.” What we are witnessing, therefore, is a gradual and significant transformation of the managerial role: various routine tasks can progressively be taken over by systems, whilst people are left to deal with ambiguous situations, exceptions, and cases where the effects of a decision cannot be fully translated into data. According to the professor, ‘those who can balance calculation and judgement will make the difference’, delegating the processing of information to the agent whilst focusing on setting priorities, interpreting the context and assessing the consequences. The issue, as one might imagine, is not limited to the relationship between a single manager and an algorithm. If the number of agents grows, a genuine digital workforce emerges that needs to be managed, and team leaders will need to know which agents are active, what they are doing, what data they are working with, what constraints they are adhering to, and what results and costs they are generating.

‘The manager will need to orchestrate and understand who does what best,’ concludes Gubitta, ‘and this means allocating tasks by entrusting agents with those that are data-intensive and require speed, comparison and operational continuity, whilst leaving experience, relationships, creativity, contextual interpretation and the management of unforeseen events to people. The further, indispensable link in the chain is managing interdependencies – that is, who initiates a task, who takes charge of the next step, when a person intervenes, and when an agent can proceed autonomously.” Paradoxically, technology can therefore bring the human factor back to the centre of leadership: if part of routine supervision is automated, the manager or team leader can devote more time to developing people, providing feedback, managing conflicts and building trust. The question every company must ask itself, Gubitta suggests in this regard, is how many people currently employed in middle management roles will be able to make this leap.

The new frontier is cognitive capital

Another major transformation currently under way – one that is less visible but set to have significant consequences for the relationship between businesses and workers – is that of professional knowledge. To function effectively, agents do not merely require large amounts of data; they must also be educated and trained by those who understand the work, so that experience, judgement criteria, processes, exceptions and routines are transferred to the algorithms. In a nutshell, this is part of that tacit knowledge that is built up over the years. The Brodolini Foundation’s research introduces, in relation to this concept, the distinction between ‘contributed capital’ and ‘captured capital’. As Maria Laura Fornaci, coordinator of the Future of Workers Observatory, explains, ‘if this transfer takes place without pre-established rules on its use, reuse and exploitation by the organisation, it risks becoming “captured capital” – that is, individual professional value absorbed by the company, via the agents, in a non-transparent manner’. Consequently, the problem goes beyond the mere fact that AI will replace a certain number of jobs: to whom will the value generated by workers’ knowledge be distributed once that knowledge has been incorporated into corporate systems?

If, on the other hand, the objectives, traceability, limits of use and forms of exploitation are defined and agreed upon in advance, that knowledge can instead become ‘contributed capital’ from which both parties can benefit. ‘The company,’ Fornaci explains in this regard, ‘provides infrastructure and computational capacity, whilst people contribute cognitive capital. But if part of that knowledge continues to generate value through the worker even after it has been transferred, it is legitimate to ask how that value should be recognised and redistributed.’ The debate on algorithmic technology thus shifts away from the issue of job displacement towards that of ownership and the distribution of the value produced by cognitive labour.

Who decides when the car makes a decision?

The issue of accountability is the final consideration. The research highlights a relatively high level of trust in AI recommendations, but also a strong sense of caution regarding the complete delegation of decision-making. In fact, 63 per cent of the managers interviewed consider human oversight and ultimate responsibility to be non-negotiable, reiterating the concept that artificial intelligence must support the decision-maker, not replace them. However, validating a decision generated by an algorithm means taking responsibility for it without necessarily having followed every step of the process, and this gives rise to the risk of a new form of ‘decision fatigue’, in which humans become the final link in a chain that they are no longer fully able to control.

This, in all likelihood, is precisely the key factor that will determine the quality of the skills transition, because agent-based AI requires (in addition to a systemic view of work to ensure the balance of the entire process) new governance skills to establish the limits of the machine’s autonomy, manage escalations, verify and ensure the traceability of decisions, and integrate algorithmic recommendations with human experience and judgement. Moreover, what is at stake is not only the relationship between humans and machines but also that between the organisation and the worker.

‘If part of that knowledge continues to generate value through the agent even after it has been transferred, it is legitimate to ask how that value should be recognised and redistributed.’ According to Fornaci’s interpretation, the issue cannot, in short, be addressed solely in economic terms; rather, rules are needed on the purpose, traceability, limits on use and exploitation of the transferred knowledge. Otherwise, there is a risk that the increase in agents’ autonomy will be perceived by workers as a loss of their professional value, fuelling resistance and conflict rather than collaboration. Hence, according to the coordinator of the Future of Workers Observatory, there is a need to redefine the relationship between capital and labour in a new way: the company provides infrastructure and computational capacity, whilst people contribute cognitive capital.

Ultimately, the challenge is not merely to decide how much work to entrust to AI, but to determine which tasks should remain with people, which responsibilities cannot be delegated, and how to redistribute the value generated by the human knowledge embedded in machines. And it is in this context that agent-based AI becomes, in every respect, a matter of organisation and management.