TLDR: AI has moved past the hype phase inside enterprises. The challenge is no longer adoption, it is execution. Costs are exploding, hybrid data estates persist, most models run on incomplete data, and security sits in a paradox of confidence and anxiety. The next era won’t be won by who experiments first, but by who masters scale.

Enterprises were promised an AI revolution. What they got instead was ballooning costs, fragmented data, and a security landscape that looks less stable the deeper you go.

The hype is still loud, but the numbers tell a quieter, more complex story.

Recent surveys of enterprise IT leaders make one thing clear: adoption is no longer the hurdle. Almost every major organisation now runs some form of AI across their operations.

The problem is scale.

Four realities stand out, and together they show why execution, not ambition, has become the defining challenge.

Pilot projects once looked cheap and impressive. Scaling them across an enterprise has turned AI into one of the biggest new cost centres. Last year, only 8% of leaders worried about compute costs. This year it is 42%. That’s a 34-point jump in a single cycle.

The irony is that this sticker shock is happening because projects are working. Over half of leaders report measurable value from AI. But those early returns are running into the hard economics of infrastructure.

Success is no longer about building proofs of concept. It is about whether organisations can afford to sustain them.

Takeaway: AI budgets will shift from experimentation to efficiency. Winners will be those who can control unit economics at scale.

Cloud-first narratives dominate headlines, but enterprises live in a hybrid world. Private and public cloud are widespread, yet almost four in ten organisations still rely on on-prem mainframes. A third continue to run distributed architectures inside their own data centres.

This is not inertia. Control, compliance, and the gravitational pull of existing data keep on-prem systems critical. For AI leaders, this creates a strategic shift: platforms must now bring intelligence to the data, not the other way around.

Takeaway: The next generation of AI platforms will be judged by their ability to operate across hybrid estates, not in a single environment.

Ninety-six percent of enterprises say AI is integrated into their business processes. But integration is not the same as completeness. Only 9% of leaders say all their data is accessible for AI. Just 38% say most of it is.

This gap matters. Models trained on partial data sets make partial decisions. For large language models in particular, an incomplete foundation means reduced accuracy, weaker ROI, and missed opportunities. The limiting factor is not model capability. It is data control.

Takeaway: The enterprises that master full data accessibility will be the ones that unlock proprietary advantage from AI.

Executives hold two positions at once. Almost half cite security and compliance risks as major barriers. Data leakage during model training and unauthorised access are the top concerns. At the same time, three-quarters say they are very confident in their organisation’s security posture.

This paradox reflects a transition. Enterprises have invested heavily in security since last year, and fewer now say it is their biggest challenge. But the attack surface is expanding faster than the fixes. Confidence and anxiety coexist because both are true.

Takeaway: Security is no longer a box to tick. It is a continuous race against the threat horizon.

Enterprise AI is not collapsing under the weight of failed pilots. It is straining under the realities of scale: cost, data access, hybrid infrastructure, and security.

The shift is clear. Enterprises don’t fail on ambition. They fail on execution.

The next era of AI will be decided not by who experiments first, but by who builds the foundations that can actually last.