Most enterprises don’t have an AI problem. They have a data foundation problem — and it’s quietly deciding who wins the AI race.
Walk into any enterprise boardroom today and you’ll hear the same confident line: “We’re investing heavily in AI.” Walk into the data center — or more accurately, the data estate — and a very different picture often emerges: fragmented pipelines, inconsistent governance, and dozens of AI pilots that never made it past the demo stage.
“The real bottleneck is not the model; it is the readiness of enterprise data, governance, and technology foundations,” Verma says. “Many companies have invested in cloud and data platforms, but their data remains fragmented, inconsistent, or poorly governed, making production AI difficult.”
The Shift From Experimentation to Execution
Over the past 18 months, Verma has watched enterprise AI cross an important threshold. The conversation inside IT organisations has changed — and so has the pressure CIOs are under.
“The question is no longer whether to adopt AI, but how to make it reliable, governed, and scalable,” he explains.
Part of what’s driving this shift is a hard realisation: the platforms enterprises built over the last decade were never designed for what AI now demands of them. “Traditional platforms were built for human interpretation, while AI agents need data that is consistent, trusted, auditable, and production ready,” Verma notes. Humans, in other words, could always work around messy data — squinting at a spreadsheet, cross-referencing a system nobody fully trusts, calling a colleague to confirm a number. AI agents can’t. They either get clean, governed, contextualized data, or they fail silently, expensively, or dangerously.
That realization is reshaping where enterprises are putting their money. Verma points to three areas absorbing the fastest-growing share of client demand: AI-ready data foundations, legacy data modernization, and agentic AI — systems that don’t just answer questions but actively automate engineering and operational work, with humans still very much in the loop. “The larger shift is from isolated pilots to enterprise-scale AI operating models,” he says.
What Separates a Pilot From a Production Deployment
If there’s one statistic that haunts CIOs right now, it’s this: most enterprises can point to dozens of AI pilots and only a handful of production deployments. Verma sees this pattern constantly — and he’s equally consistent about the diagnosis.
“The real barrier to scaling AI is rarely the model itself,” he says. “More often, it is the lack of a trusted, governed, and scalable enterprise data foundation.”
He offers two illustrative examples, both anonymized. A global airline used Xebia’s Axis platform to modernize fragmented operational and customer data before attempting to scale AI — pairing AI agents with human data engineers to automate estate discovery, code conversion, validation, and migration, all while preserving governance and quality controls. The result: a migration completed nearly three times faster than conventional approaches, and a genuinely production-ready data platform underneath it.
In a separate case, a global retailer consolidated duplicate data pipelines into a single governed platform with continuous monitoring — a less glamorous story than a flashy generative AI use case, but one that materially improved consistency and trust across reporting and operations.
The common thread, Verma says, isn’t a clever model or a novel algorithm. “Organisations that scale AI first fix governance, data quality, platform modernisation, and ownership. AI agents then automate repetitive engineering work while human experts provide oversight.” Scaling AI, in his framing, is less an act of innovation than an act of discipline. “The difference between pilots and production is discipline, with AI treated as a transformation program rather than a technology experiment.”
Cutting Through the ROI Noise
Every CIO today is fielding some version of the same question from the board: where’s the return? And every CIO knows that not every AI use case can honestly answer it yet.
Verma is direct about where the real returns are showing up — and where the hype has outrun the results. “AI ROI comes down to whether it solves a real business problem. The strongest returns come when enterprises start with the outcome, not the technology.”
The use cases delivering measurable value, he says, tend to be the unglamorous ones embedded deep in core operations: data platform modernization, automated migration, data engineering and quality, governance, and real-time decision-making. Through Xebia Axis, for instance, AI agents work alongside data engineers to assess, migrate, monitor, and operate enterprise data platforms faster — without loosening the grip on governance and compliance. AI-assisted data operations are showing similarly concrete payoffs, Verma notes, through continuous monitoring, root-cause analysis, and proactive issue detection that improve data reliability across the board.
On the other side of the ledger sit the pilots that generate headlines but not outcomes. “The use cases generating more hype than results are pilots without a business owner, production roadmap, trusted data, or governance,” Verma says. “These may demonstrate capability but rarely scale or prove sustained ROI.”
His conclusion is blunt: the enterprises actually winning with AI aren’t the ones with the flashiest demos. They’re the ones treating AI as an operating model — governed data, modern platforms, clear ownership, and human expertise working in concert to deliver value that repeats, quarter after quarter.
A Readiness Checklist for CIOs
So how does a CIO know, walking into an AI initiative, whether the foundation underneath it can actually hold the weight? Verma lays out a sequence that starts, notably, nowhere near the technology stack.
“The first test is not the technology stack; it is the business objective,” he says. “CIOs should begin by asking what value AI is expected to create, which use cases matter most, and which workloads deserve priority. Architecture should follow those answers, not lead them.”
Only after that comes the harder, less comfortable diagnostic: data readiness. “AI needs data that is trusted, governed, accessible, and production-ready,” Verma says. “Yet in many enterprises, data is still scattered across systems, documents, emails, and repositories, with uneven quality and fragmented governance.” It’s a scenario most CIOs will recognize immediately — and one that humans have quietly compensated for, for years, without anyone quite realizing how much manual judgment was papering over the cracks. AI agents offer no such patience. “Humans can work around this complexity; AI agents cannot.”
Then there’s governance — not as a compliance checkbox, but as infrastructure in its own right. “Governance is equally critical,” Verma says. “Clear ownership, access controls, audit trails, data quality processes, residency rules, and bias controls must be built into the data lifecycle from the start.”
His practical advice to CIOs before they commit further budget to AI is refreshingly unglamorous: assess the data estate, modernise fragmented platforms, remove duplicate pipelines, and put governance and observability in place — before, not after, the AI investment.
The Real Lesson
Strip away the acronyms and the platform names, and Verma’s message amounts to a quiet correction of the industry’s instincts. Enterprises have spent 18 months racing to adopt AI as if the technology itself were the scarce resource. It isn’t. Models are increasingly commoditised, accessible, and interchangeable. What’s scarce — and what’s proving decisive — is the unglamorous, unfinished work of making enterprise data trustworthy enough for a machine to act on without supervision.
“Most organisations do not lack data,” Verma says. “They lack the right data in the right shape. That is what separates scalable AI from endless experimentation.”
For CIOs still measuring AI progress by the number of pilots underway, that’s a hard but clarifying reframe. The enterprises that will define the next phase of AI adoption won’t be the ones that moved fastest into experimentation — they’ll be the ones patient enough to build the foundation first, and disciplined enough to treat AI not as a project with a finish line, but as an operating model built to last.