Companies of all sizes are accelerating the adoption of artificial intelligence agents, attracted by the promise of increased efficiency and reduced operating costs. But this race hides a little-discussed risk: a significant portion of these projects are not expected to survive beyond the next two years, consuming investment without delivering the promised return.
The warning is supported by data from two of the most respected institutes in the technology sector. Gartner projects that 40% of enterprise applications will have artificial intelligence agents focused on specific tasks by the end of 2026, compared to less than 5% in 2025, and at the same time predicts that more than 40% of agentic AI projects will be canceled by the end of 2027 due to rising costs, unclear business value, or inadequate risk controls. A Deloitte study reinforces the scenario from another angle: almost three-quarters of companies plan to adopt agentic AI in the next two years, but only 21% of them claim to have a mature governance model for these agents.
According to Carlos Pisani, a systems architecture specialist and founding partner of ArcH, the numbers describe a pattern he recognizes in practice, long before any report confirms it. “The problem is almost never the AI model chosen. It's placing an autonomous agent on top of a system that never had clear technical specifications or a validated architecture. Each new layer added to this base increases the technical debt instead of solving anything,” he states.
According to Pisani, there is a change in the behavior of companies that has not yet reached the way they build software. “The market has matured in the way it thinks about business. The demand for operational efficiency and financial sustainability has grown; this is no longer up for debate. The problem is that this maturity has not reached the development of systems. Many companies have evolved their business model and continue to develop software in the old way, without specification and without architecture validation. That's where the technical debt is born, which then hinders precisely the efficiency that the business itself has come to demand.”
According to Pisani's analysis, the Deloitte data confirms the diagnosis through the lens of governance. "The fact that only 21% of companies have a mature agent governance model is not a compliance problem, it's a symptom of weak architecture. Agent governance presupposes that there is something structured to govern. When the system is born without specification, no governance framework can solve it later."
To reduce the risk of being among the 40% of canceled projects, Pisani argues that technical specifications should be treated as business decisions, not as technical steps to be skipped. “Before asking which AI agent to use, the right question is whether the underlying architecture can handle that decision. This is what we call Spec-Driven Architecture (SDA): no technical decision is made before there is clarity about what needs to be solved and how it will be sustained over time. Companies that skip this step are not saving time; they are postponing the cancellation of their own project.”
