Eighty-five percent of organizations are actively experimenting with AI agents, yet only 5% have successfully moved them into production. According to a Cisco survey of major enterprise customers, this gap is not an anomaly. Research from Anaconda and Forrester indicates that 88% of enterprise AI agent pilots fail to reach production, a trend corroborated by findings from a16z and the MIT Sloan CIO panel. Even looking ahead, the outlook remains cautious; Gartner has projected that over 40% of agentic AI projects could be canceled by the end of 2027 due to a combination of escalating costs, unclear business value, and inadequate risk controls. Meanwhile, S&P Global Market Intelligence reported in Q1 2026 that only 31% of organizations have at least one AI agent running in production.
For many leaders, the instinct is to assume that these failures stem from a lack of capability. We are conditioned to believe that if an agent isn’t working, it simply isn’t “smart” enough yet. However, this assumption is increasingly being challenged by those building the technology. As Bryan Silverthorn, Director of AGI Autonomy at Amazon, noted at VB Transform 2026, “The question is not whether AI agents are capable enough. It’s whether they are reliable enough to trust with real business processes.”
The reality is that agents often perform exceptionally well in controlled, internal evaluations, only to collapse when exposed to the variability of real customers. The problem is not the model’s intelligence, but our ability to measure and manage its behavior. To diagnose this, researchers at Princeton have proposed a four-dimensional reliability framework. They define reliability through consistency (repeatable outcomes under nominal conditions), robustness (graceful degradation under perturbations), predictability (confidence aligned with accuracy, and the ability to defer under uncertainty), and safety (ensuring bounded harm even when failures occur). Their findings suggest that simply pushing for more capability — making the model “smarter” — has yielded only small improvements in these core reliability metrics.
This shift in focus changes what is required of enterprise managers and workers. Silverthorn offers a useful metaphor: treat agents like interns. They are, in his words, “powerful but occasionally clueless, capable of amazing work and spectacular derailment.” This perspective shifts the burden from pure software engineering to management. If an agent is an intern, it requires oversight, clear boundaries, and a system for handling its inevitable mistakes. The Princeton HAL Reliability Dashboard provides a way to evaluate agents across these dimensions, moving the conversation away from abstract performance benchmarks toward operational readiness.
The operational implication is that the organizations most likely to escape the 85% pilot ceiling are not necessarily those with the most advanced models, but those with the best agent managers. This requires a fundamental change in how IT teams approach deployment. Instead of asking what an agent can do, managers must ask how the agent behaves when it encounters something it doesn’t understand, and whether the business process can survive the agent’s “spectacular derailment.”
Moving beyond the pilot phase requires us to stop treating these tools as autonomous, perfect systems and start managing them as high-potential but fallible contributors. We are currently in a phase where the technology is genuinely capable, but the operational framework to support it is still maturing. Crossing the gap from pilot to production will not be solved by a breakthrough in model intelligence alone. It will be solved by the rigorous application of reliability standards and the development of management practices that treat these agents with the same oversight we apply to any other valuable team member. The question is no longer whether agents will eventually work, but whether enterprises can build the management discipline required to make them reliable enough to trust.