Software engineering and the change underway

The market is saturated with miraculous narratives about AI: instant accelerations, demos that seem to solve complex problems in an instant. The reality in companies is more moderate. Gartner currently places GenAI in the "trough of disillusionment" and estimates that by 2028 at least half of the projects will exceed budget costs due to poor architectural choices and lack of operational know-how. According to MIT, 95% of generative AI pilot projects have not generated tangible economic returns, against an estimated expenditure of between 30 and 40 billion dollars. It's not surprising that many C-levels feel torn between enthusiasm and fear of being left behind: in IBM's CEO Study 2025, 64% of CEOs admit to investing more out of fear than from a real understanding of the value. The answer is neither blind trust nor principle-based skepticism: it is a method that distinguishes what generates value in production from what only works in a presentation.

An application has never been worth the days spent creating it, but for the importance of the problem it solves. For decades, time has been a convenient shortcut for estimating costs and prices; today, with AI compressing development times, that approximation no longer holds. Clients' questions have already changed: no longer "how many days are needed?", but "which problem are we solving, which indicator is improving, how do we move from experiment to industrial scale?". It is the transition from an effort logic to an outcome logic: an operating cost that goes down, a time-to-market that shortens, a risk that is neutralized. It is a transition that redefines not only pricing models, but the very sense of the relationship between those who create software and those who commission it: those seeking excellence can choose partners capable of linking work to measurable objectives and taking responsibility for the result.

The tension between autonomy and dependence manifests on two levels. On the organizational level, AI makes activities more accessible that previously required broader development capabilities, and can lead to internalizing software production. But producing more code does not mean being more autonomous: DORA research shows that AI amplifies existing capabilities and weaknesses, and the benefits of speed are absorbed by bottlenecks in testing, security, and deployment when the foundations are inadequate. On the technological level, an apparent paradox emerges: the cost of a single inference drops rapidly, but Agentic AI multiplies the number of inferences needed to achieve a result, and independence does not simply coincide with running a model locally, because every deployment remains embedded in an interdependent ecosystem of hardware, infrastructure, data, and expertise. On both levels, autonomy is not about owning every component, but about consciously choosing and managing one's dependencies, ensuring the right skills, interoperability, portability, and the possibility to replace partners and technologies.

The agency era transforms the profession of software builders: mastery shifts from writing code to the ability to instruct, coordinate, and monitor systems operating autonomously. Faros AI telemetry on 22,000 developers measures the shift: 66% more completed epics per person, but also 54% more bugs per developer and code review times increased fivefold. The true differential is not the machine, but the human interpreter who ensures the governance of quality, risk, and consistency of the outcome. The same applies to relationships: Gartner predicts that over 40% of agentic AI projects will be canceled by 2027 due to rising costs, uncertain benefits, or inadequate risk controls. In high-budget and high-risk decisions, decision-makers seek not just technology: they seek presence, listening, continuity. Change governance itself becomes part of the service.