New Delhi, May 29 -- Why enterprises are still stuck in AI's 'Pilot Purgatory'

The conversation in enterprise boardrooms has shifted. For the better part of a decade, technology leaders were preoccupied with a singular anxiety: would AI actually work? That question has been answered, emphatically and repeatedly, in the affirmative. The question that now haunts CIOs, strategy committees and investment committees is considerably more uncomfortable: if the technology works, why are so few organisations managing to extract value from it?

What the industry is living through, in the unglamorous language of operational research, is a scaling problem. AI's capabilities have outpaced enterprise readiness to deploy them at a meaningful scale. The result is a peculiar graveyard of successful pilots, initiatives that demonstrated genuine intelligence in controlled conditions, only to stall when confronted with the full complexity of enterprise operations. According to MIT's The GenAI Divide: State of AI in Business 2025, nearly 95% of AI initiatives still fail to reach enterprise-scale value. Enterprise AI, it is now clear, has entered its accountability phase.

Infrastructure enthusiasm is outpacing execution readiness. One of the more costly misapprehensions of the AI boom has been the belief that infrastructure investment is equivalent to operational readiness. It is not. Across industries, organisations have committed to large-scale compute architectures and GPU clusters before establishing what, precisely, those architectures are meant to accomplish, and for whom.

The consequences are predictable in retrospect. Expensive deployment models lock organisations in before value has been demonstrated. Redesign cycles follow. Downstream inefficiencies multiply. Arun Shetty, Chief Technology Officer at Cisco India and South Asia, identified the underlying dynamic with characteristic bluntness: starting is not the hard part, scaling is. The gaps, he notes, run across compute, networking and data architectures alike.

The data bears this out. As per 2024 State of AI Infrastructure Survey, Nearly 74% of companies remain dissatisfied with their current GPU scheduling tools. That a technology supposedly transforming global business cannot yet be reliably scheduled is revealing. Pilot success, it turns out, has been systematically masking production failure. The larger constraint is organisational, not technological

Infrastructure, though, is only the first layer of the problem. Beneath it lies a more stubborn constraint: organisational unreadiness. Governance fragmentation, unclear ownership structures and the persistent misalignment between AI capabilities and actual workflows are preventing initiatives from ever graduating beyond experimentation.

Rishi Aurora, Managing Partner at IBM Consulting India and South Asia, frames it with an observation that has become something of a governing truth for anyone working seriously in enterprise AI: the biggest constraint is not the AI itself, it is enterprise readiness. Technical debt, fragmented data ecosystems, legacy operating models and talent shortages recur as barriers. More critically, many AI initiatives remain disconnected from the operational systems and governance structures that would give them staying power, turning them into expensive experiments rather than genuine transformation.

The governance picture is particularly sobering. According to Deloitte's Global AI Governance Survey 2026, only 39% of chief executives believe they currently have reliable AI governance structures in place. That figure in 2026, after years of public conversation about responsible AI speaks not to a lack of awareness but to a structural failure to translate intent into institutional architecture.

The Invisible AI Imperative

There is an irony embedded in the phrase "AI transformation." The organisations achieving the most durable results are not, by most appearances, being transformed by AI at all. The technology is simply invisibly making their operations more effective. This is the logic behind what practitioners are calling workflow-native intelligence: AI embedded directly into operational systems rather than deployed as a layer of standalone applications that employees must consciously choose to use. The distinction matters more than it might appear. AI that sits outside core workflows is, by its nature, optional. Adoption plateaus. Business impact stalls.

DBS Bank's approach in India illustrates the alternative. Chief Technology Officer Ramesh Mallya describes how the bank has focused on integrating AI directly into customer journeys and operational systems, not treating it as a separate technology investment measured in isolation. The emphasis is on incremental modernisation, embedding AI into existing processes, and maintaining human oversight throughout. The goal is for AI to deliver its greatest value precisely when it ceases to be visible as AI at all.

The Accountability Reckoning

The financial stakes of continued underperformance are becoming impossible to ignore. According to the IBM CEO Study 2025, only 25% of AI initiatives are currently delivering their expected return on investment, and barely 16% have scaled enterprise-wide. McKinsey's State of AI 2025 found that only 39% of organisations report measurable EBIT impact from their AI programmes, meaning more than 60% still see no tangible effect at the level of enterprise earnings.

These are not numbers that boards can indefinitely overlook. The "we are learning" rationale that sustained AI experimentation budgets through the early part of this decade is exhausted. What boards now demand is accountability: measurable outcomes, institutional governance and a credible path from pilot to production.

CIOs have registered the shift. While 57% of enterprise leaders surveyed identified AI as a strategic priority, only 6% had managed to scale initiatives beyond the pilot stage into genuine enterprise deployment. The gap between aspiration and execution has rarely been so empirically documented.

The New Competitive Divide

What separates the organisations that will emerge from this period with a durable AI advantage from those that will have spent the decade on expensive pilots is, in the end, a question of operational discipline.

The next generation of enterprise AI leaders will not be defined by the ambition of their infrastructure investments or the sophistication of their models. They will be defined by their capacity to build integrated AI operating systems, environments in which infrastructure, governance, workflow integration and outcome measurement function as a continuous feedback loop rather than as separate departmental initiatives.

High-performing organisations, McKinsey's research suggests, continuously optimise their AI systems using adoption metrics, user feedback and operational outcomes, not pilot success metrics, which measure something altogether different.

Enterprise AI is no longer a technology problem. It is a management problem. And management problems, unlike model capability problems, admit of no shortcut. The organisations that will lead are not those that experimented the fastest, they are those that scaled with the greatest discipline, the clearest governance and the most uncompromising focus on measurable value.

One proven mechanism for closing that gap is institutionalising scale via an AI Programme Management Office. An AI PMO bridges the divide by owning portfolio oversight, standards and accountability for enterprise-wide AI value, transforming scattered initiatives into a governed, outcome-driven programme.

Published by HT Digital Content Services with permission from TechCircle.