From transformation to delegation: setting boundaries, context and accountability when AI acts on business processes

Objectives before tools 

AI is a means, and it belongs inside the organization’s objectives. Value comes from redesigning processes, and the risk is pedaling an exercise bike: spending on licenses and consumption without getting anywhere. Istat data (Italy’s national statistics institute, Enterprises and ICT 2025) show adoption is still uneven: 16.4% of companies with at least ten employees use at least one AI technology, up from 8.2% in 2024, with 53.1% among large companies and 15.7% among SMEs. Adopting department by department scatters resources: the company has to be redesigned as a coherent whole. 

Agents, context and data 

An agent detects events, reasons, makes recommendations and takes actions without human intervention at every step. People retain decision-making, governance and control. If a supplier reports a two-month delay, the agent assesses the effects on production, deliveries and cash before the meeting, so time goes into deciding rather than reconstructing the facts. An assistant can also work from intent: you state an outcome, such as shortening payment collection times, and the system analyzes, recommends and carries out some actions. Agents go wrong when context is missing. Context has two parts: processes and rules, which live in the systems, and people’s know-how, which was never written down. The second weighs more in small companies, where the instinct for when to call a customer does not sit in the CRM. That is why the reasoning behind decisions needs documenting too. Data quality matters as well, including semantic correctness, and it has to be validated by someone who knows the domain. According to Istat, among companies that evaluated AI without adopting it, nearly six in ten cite a lack of skills and just under half cite poor data quality. 

Channel, sovereignty, security 

Context travels through a channel, and the channel has to be designed. People who work with AI take a lot of context for granted: a program works and has security problems because security was never asked for. Early machine tools took over a century to get effective interfaces, and AI follows a similar pattern. The channel must run both ways: people share knowledge more willingly when they receive some in return. 

Sovereignty is about where know-how stays. Large providers are moving into companies, and the risk is giving away a competitive advantage that then becomes available to everyone indiscriminately. On employment matters, Law No. 132/2025 provides for informing workers about the use of AI (art. 11) and for an Observatory at the Ministry of Labour (art. 12). On security, the 2025 Report of Italy’s National Cybersecurity Agency (ACN) records 2,729 cyber events handled by CSIRT Italia, 38% more than in 2024, and 615 incidents with confirmed impact, up 7%: more electronics on critical assets widens the surface to protect. 

Costs and suppliers 

With agents, cost adds the license to resource consumption, and it is hard to forecast when several agents work across different platforms: it has to be designed together with the architecture. Companies choose between the agents of platforms they already use and an architecture of their own, and the two routes will probably coexist. The relationship with suppliers changes too: if work is no longer only human, the man-day loses meaning as a unit of measure, and a discount on the rate does not address the problem. 

The models still missing 

AI will take over basic tasks, and people’s role will become more relevant, because their intelligence will be deployed at the critical points of decisions. Shared models of knowledge, work, accountability, control and human in the loop are missing. No organization can receive them from outside, and it makes sense to build them together. Two questions remain: what objective justifies the use of these systems, and who answers for the outcome.