A significant portion of enterprise AI expenditures remains within the research and development phase.

      Recent research quantifies something many IT leaders have already suspected: the majority of AI pilots never transition into production, and the associated spending isn’t disappearing; it’s in a state of limbo. Jon Bitz, the chief relationship officer and co-founder at KloudStax, believes that this limbo isn’t as much of a failure as it appears from the outside.

      The statistics consistently point to the same conclusion, regardless of who is measuring. Forrester’s recent analysis on agentic AI found that 75% of enterprise leaders claim to be adopting it, yet only a small fraction have it operating beyond limited pilots. Gartner forecasts that by 2026, organizations will abandon 60% of AI projects that lack AI-ready data and integration infrastructure. Additionally, Deloitte’s enterprise survey revealed that over one-third of companies are still utilizing AI at a superficial level, showing minimal real change in their day-to-day operations.

      When these findings are combined, a trend emerges: enterprises are approving AI budgets more swiftly than they are turning that expenditure into anything like a stable, operational system. The gap between these two aspects has become a pressing question in enterprise cloud discussions.

      The spending isn’t wasted; it’s just parked. Jon Bitz, who is deeply involved in budget discussions as a Google Cloud partner, does not view the situation as dire as the abandonment statistics suggest. “A significant portion of AI cloud spending is still occurring in pre-production, and that’s not necessarily a negative thing,” Bitz remarked. “Testing, experimentation, and validation are all crucial for properly adopting AI. This phase is essential for determining what works and what doesn’t.”

      However, he notes a genuine concern regarding the prolonged duration of this phase without intervention to advance it. “In many instances, a significant amount of funded projects remain outside stable, repeatable production workflows,” he stated.

      Bitz believes the solution is not to reduce the experimentation budget, which is often the instinct of finance teams when pilot results become uncomfortable. Instead, organizations should focus on systematically converting insights gained during experimentation into operational systems. “Our aim is not necessarily to cut experimentation but to transition it,” he explained. “We should take the insights gained and turn them into production systems, shifting spending from testing to workloads that genuinely generate value for the business.”

      When asked where cloud spending is currently generating value, Bitz mentions areas less sensational than many of the recent AI headlines. Not related to new model releases or larger context windows, but to the fundamental tasks that were already essential to running a business prior to the advent of generative AI. “True value is seen in core workflow automation – transforming manual, labor-intensive processes into production-ready systems,” he noted, citing support operations, engineering productivity, and AI embedded directly into revenue-generating workflows as key areas of impact.

      In his experience, misallocation often follows a familiar pattern: teams tend to pursue better models or deploy more GPUs since that’s the more straightforward part, while overlooking the tougher work of structuring data and defining workflows. “Without structured data and clearly defined workflows, moving anything into production becomes extremely challenging,” Bitz remarked.

      There’s also been increasing dialogue recently regarding whether enterprises are beginning to challenge hyperscaler pricing, especially as AI workloads complicate monthly billing. According to Flexera’s 2026 State of the Cloud data, cost unpredictability related to dynamic AI workloads has surfaced as one of the primary challenges enterprises face, with nearly half of large organizations now maintaining a dedicated AI governance function to monitor this. However, Bitz indicates that this doesn’t necessarily translate into pushback against the platforms themselves based on his observations.

      “We are witnessing more scrutiny, but not significant pushback on the hyperscalers or the value of their solutions and platforms,” he commented. “The focus is more on how spending is structured and managed.” The questions organizations are raising tend to be more about specific drivers of costs rather than an overall critique of cloud pricing: what is actually contributing to the bills, what relates to production versus pilots and proofs of concept, and where there might be opportunities for optimization.

      This scrutiny is encouraging companies to lean towards more opinionated architectural choices, adopting serverless and managed services, and aligning workloads more closely with their actual needs, alongside stricter governance regarding who can initiate what. Partners are also taking on more of this optimization work. “It’s not about challenging the price so much as figuring out how to use these platforms effectively,” Bitz said. “When the architecture, data, and usage patterns are correct, the economics tend to follow.”

      It can be tempting to view escalating AI cloud costs solely as a computing issue, since that’s the most visible item in the budget. However, Bitz argues that this is rarely the core issue. “Compute is an easy target, but inefficiency usually reflects gaps in data, architecture, governance, and undefined workflows,” he explained. Simply adding a more extensive model or additional GPUs to a system that was never designed for AI