Most enterprise AI spend still hasn’t left the lab

New research puts a hard number on something a lot of IT leaders already suspected. Most AI pilots never make it into production, and the spend behind them isn’t disappearing, it’s just sitting in limbo. Jon Bitz, chief relationship officer and co-founder at KloudStax, says that limbo isn’t the failure it looks like from the outside.

The numbers keep landing in the same neighborhood no matter who’s counting. Forrester’s latest research on agentic AI found that three-quarters of enterprise leaders say they’re adopting it, but only a small minority have it running in anything beyond limited pilots. Gartner predicts that through 2026, organizations will abandon 60 percent of AI projects that aren’t backed by AI-ready data and integration infrastructure. And Deloitte’s own enterprise survey found that more than a third of companies are still using AI at a surface level, with little real change to how they actually work day to day.

Put those together and a pattern starts to form. Enterprises are approving AI budgets faster than they’re converting that spend into anything resembling a repeatable, working system, and the gap between the two has become one of the more uncomfortable open questions in enterprise cloud right now.

The spend isn’t wasted, it’s parked

Jon Bitz, who spends most of his time inside these budget conversations as a Google Cloud partner, doesn’t think the story is as bleak as the abandonment numbers make it sound.

“A meaningful portion of AI cloud spend is still happening in pre-production, and that’s not a bad thing,” Bitz said. “Testing, experimentation, and validation are all critical parts of adopting AI correctly. You need that phase to understand what works and what doesn’t.”

Where he does see a real problem is how long that phase tends to drag on without anyone stepping in to move it along. “In some cases, a large amount of what’s being funded hasn’t made its way into a stable, repeatable production workflow yet,” he said.

His read on the fix isn’t to cut the experimentation budget, which is the instinct a lot of finance teams reach for once the pilot numbers get uncomfortable. It’s to get more deliberate about converting what the experimentation phase actually teaches an organization into something that ships. “Our goal isn’t necessarily to reduce experimentation, it’s to convert it,” Bitz said. “To take what’s being learned and turn it into production systems, so spend shifts from testing into workloads that are actually driving value inside the business.”

Where the money actually goes to work

Ask Bitz where cloud spend is already paying off and he points somewhere less flashy than most of the AI headlines from the last year. Not new model releases, not bigger context windows, just the unglamorous stuff that was already running a company before generative AI showed up.

“Real value is showing up in core workflow automation, taking manual, time-intensive processes and turning them into production-ready systems,” he said, pointing to support operations, engineering output, and AI embedded directly into revenue-generating workflows as the places where it’s landing.

The misallocation, in his experience, tends to follow a familiar script. Teams chase a better model or spin up more GPUs because that part is easy, while the harder work of structuring data and defining workflows gets skipped. “If the data isn’t structured and workflows aren’t clearly defined, it becomes very difficult to move anything into production,” Bitz said.

Hyperscaler bills are getting more scrutiny, not less trust

There’s also been a louder conversation lately about whether enterprises are starting to push back on hyperscaler pricing itself, especially as AI workloads make monthly bills harder to predict. Flexera’s 2026 State of the Cloud data backs that up from the finance side, with cost unpredictability tied to dynamic AI workloads showing up as one of the top hurdles enterprises report, and nearly half of large organizations now running a dedicated AI governance function to keep it in check. Bitz says that’s not quite adding up to pushback on the platforms themselves in what he’s hearing in practice.

“We’re starting to see more scrutiny, but not so much pushback on the hyperscalers themselves or the value of the solutions and platforms,” he said. “It’s more about how spend is structured and managed.” The questions companies are actually asking tend to be narrower than a referendum on cloud pricing: what’s actually driving the bill, what’s tied to production versus pilots and proofs of concept, and where there’s room to optimize.

That scrutiny is nudging companies toward more opinionated architecture choices, serverless and managed services and workloads sized to what they actually need, along with tighter governance over who can spin up what. Partners are picking up more of that optimization work too. “It’s less about challenging the price and more about asking how to use these platforms correctly,” Bitz said. “When the architecture, data, and usage patterns are right, the economics tend to follow.”

Inefficiency has a paper trail, and it rarely starts with compute

It’s tempting to treat runaway AI cloud bills as a compute problem, since that’s the line item everyone can see. Bitz says that’s almost never where the real story is.

“Compute is an easy thing to point to, but in most cases, inefficient spend is really a reflection of gaps in data, architecture, governance, and undefined workflows,” he said. Throwing a bigger model or more GPUs at a system that was never designed to support AI in the first place doesn’t fix the underlying mess, it just makes the same inconsistency more expensive to run. As Bitz put it, AI isn’t creating new problems inside an organization so much as it’s amplifying the ones that were already there.

The gap between big budgets and small ones isn’t really about budget

There’s a separate worry running through a lot of mid-market conversations right now, which is that cloud pricing and infrastructure complexity are quietly widening the gap between companies that can afford to experiment with AI and companies that can’t. Bitz thinks the gap is real but that people are pointing at the wrong cause.

“Complexity plays a bigger role than just a budget gap between SMB, mid-market, and enterprise, although the barrier to entry is real,” he said. Larger organizations have room to absorb a few expensive misses along the way. Smaller ones don’t get that same margin for error, so they end up needing to be more precise from the start rather than less ambitious. “The companies we see winning with AI aren’t necessarily the ones spending the most, they’re the ones implementing the cleanest systems around well-defined use cases,” Bitz said.

What happens over the next year and a half

Bitz doesn’t expect AI spend to slow down, but he does think the conversation inside enterprises is already shifting toward accountability rather than raw scale. He’s watching for organizations moving proof-of-concept work onto real production-level KPIs, more attention paid to unit economics like cost per output or per workflow, engineering teams getting measured on the AI-driven output they produce rather than how much of the tooling they’ve adopted, and more C-level visibility into what AI spend is actually buying.

“AI adoption is still early for a lot of organizations, and as teams get more comfortable, the focus naturally shifts from ‘are we using AI?’ to ‘is this actually working and driving value?‘” Bitz said. Over the next 12 to 18 months, he expects that shift to keep building, with spend continuing to grow but under a lot more pressure to show its work.