The AI budget that hurts most is not the one that gets lost, but the one that never reaches where it matters
More than $1.5 trillion in enterprise software valuations evaporated over the last two years. Not for lack of investment in artificial intelligence, but because that investment landed in the wrong place. This is the paradox that defines the current moment: companies have never spent so much on AI and, at the same time, it has never been so difficult to show where the value actually is.
Rohit Kedia, CEO of Xoriant, articulated this with precision in a recent analysis: the vast majority of corporate AI budget accumulates at the model layer — that is, in platform licenses, computing infrastructure, vendor partnerships, and proof-of-concept development. That layer captures attention, generates announcements, and produces demonstrations. What it does not produce, consistently, are results that show up on the income statement.
Gartner estimated in February 2025 that, through 2026, 60% of AI projects will be abandoned due to a lack of data that is ready to be processed. This is not a technological failure. It is a failure of budgetary architecture: the model was funded, not the infrastructure that sustains it.
The question that technology leaders should be asking themselves is not whether to invest in AI. That decision has already been made. The question is whether the money is building operational capacity or financing the appearance of modernity.
The expensive theatre no one wants to name
There is a pattern that repeats with striking regularity in boardrooms. Every week brings a new announcement of a partnership with a model provider, a demonstration of autonomous agents, a promise of transformed workflows. The noise is loud. The theatre, convincing.
When one looks beyond the press releases and asks what actually changed in any concrete way in how the company creates value for its customers, the honest answer is usually disappointing. McKinsey reported in its State of AI 2025 that only one third of companies have begun to scale their AI programs at the enterprise level, even though 88% report active investment. The rest is spending. Just not where it counts.
This phenomenon has an identifiable structure. AI budgets in 2023 and 2024 lived primarily in innovation or R&D lines, with lax return requirements and light financial oversight. That created the perfect conditions for the proliferation of overlapping tools, departmental pilots disconnected from core processes, and infrastructure contracts that no one measured against a specific business outcome.
The problem is not that companies experiment. Experimentation has value. The problem is when the experiment becomes the product, when the demonstration replaces deployment, and when the innovation budget functions as a way of appearing modern without committing to anything concrete.
Deloitte found that around 66% of organisations that adopted enterprise AI report improvements in productivity and efficiency as the main benefit obtained. That is a reasonable number. But it must be read carefully: productivity and efficiency are process metrics, not necessarily indicators of structural economic impact. Reducing the time it takes an analyst to prepare a report is not the same as reconfiguring the chain of decisions that makes that report matter.
The distinction between "using AI" and "applying intelligence" is, at its core, a budgetary distinction. Using AI means placing tools on top of existing workflows: a copilot here, a chatbot there, an analytics layer over a dashboard that already existed. Applying intelligence means something categorically different: embedding automated decision-making capacity within the way the company produces and delivers value, with traceability back to the outcome that decision affects.
That second option requires funding things that do not generate headlines: data cleansing, process redesign, modernisation of legacy architectures, team formation. Goldman Sachs noted in its March 2026 report that AI is not eating the software market; it is expanding it, because it reduces the cost of writing code while raising the ceiling of what that code can do. That implies the space of applicable value has grown. But capturing it requires having built the foundations that support it.
The four foundations that the budget ignores
There are four areas where investment systematically arrives late or falls short, and all four determine whether any spending on models produces value or simply produces activity.
Processes are the first point of failure. AI applied on top of a broken workflow produces broken results faster. Every dollar invested in the model without redesigning the process surrounding it is a dollar that generates speed, not direction. The most frequent mistake in enterprise AI programmes is to assume that the intelligence of the model will compensate for the dysfunction of the process. It does not. It amplifies it.
Technological architecture is the second problem. Fragmented legacy systems cannot support intelligence embedded at the point of decision. Purchasing a more powerful model does not resolve an integration problem. What appears to be an AI capability problem is, frequently, an unresolved technical debt problem that the AI budget never touched, because technical debt does not generate attractive demonstrations.
Skills occupy the third position and are perhaps the most costly deficit due to their invisibility. There is a difference between a workforce that knows what AI is and one that knows how to work with it. The first can answer an adoption survey. The second can redefine how a team in operations, risk, or customer service makes decisions. The transformation of capabilities remains one of the most consistently underestimated budget lines in enterprise AI programmes, treated as change management at the end of a project rather than as a delivery condition from the outset.
Data closes the picture. No model, however sophisticated, produces reliable intelligence from unreliable data. And yet data preparation receives a fraction of the investment directed toward model selection and platform acquisition. Gartner's finding is not merely statistical: it is a diagnosis of priorities. Companies invest where there is visibility and recognition. Clean, well-governed data that is connected to the right processes does not generate spectacular demonstrations. It generates results. That difference explains why 60% of projects do not survive.
A cost study published in 2026 estimates that production-ready AI systems — with genuine regulatory compliance and real scalability — cost between $250,000 and more than one million dollars per system, once engineering, data work, governance, and integration are fully accounted for. That figure includes the recurring costs of model maintenance, retraining, and monitoring. Almost no pilot was designed to sustain that structure. Which explains why pilots do not scale.
The budgetary architecture that separates those who capture value from those who merely observe it
The difference between the companies that are capturing sustainable value with AI and those that are accumulating implementation debt is not found in which models they chose. It lies in how they built the investment architecture around those models.
Companies that are generating measurable returns share a three-layer pattern. The first is foundational investment: data preparation work, process redesign, modernisation of legacy systems, and training programmes with adoption metrics. It is the unglamorous work that determines whether everything else functions. The second layer is intelligence deployment: AI integrated natively into real workflows, not as a parallel tool but as a capability within the platform, the product, or the customer journey, with direct traceability to a business outcome. The third is orchestration — both human and agentic — but it only holds value when the two previous layers have already been built.
Deloitte's projection indicates that the number of companies with more than 40% of their AI projects in production will double in the next planning cycle. That number matters less as an indicator of adoption than as a signal of what type of company will be able to compete on a structurally different cost basis.
Chief financial officers are beginning to move AI budgets from innovation lines with lax oversight into operational technology budgets, subject to the same demands applied to an investment in ERP systems or customer relationship management platforms. That has two immediate consequences. The first is that projects unable to demonstrate a measurable operational return will lose funding. The second is that the vendors and systems integrators that survive will be those capable of connecting their offering to concrete process metrics, not to promises of abstract transformation.
The core argument of Kedia's analysis — and the one that most unsettles technology teams — is that the AI investment that will matter most in the coming year is the one that today appears least attractive. That is not rhetorical paradox. It is a precise description of how value is distributed in markets with high information asymmetry: those who invest in what cannot be demonstrated in a presentation capture the value that others merely describe in their annual reports.
The budgetary architecture that builds applied intelligence is, by definition, less visible than the one that funds experiments with advanced models. But it is the only one that produces results capable of withstanding a genuine value audit. And in an environment where boards of directors are beginning to demand exactly that, the visibility of spending has ceased to be its principal advantage.
