The Economics of AI Adoption Are Backward
Most enterprise AI spending pays for the easy half of the problem and leaves the hard half unfunded.
By Dr. Reza Olfati-Saber, Founder & Chief Scientist, Wisdom Agent, Inc.
The Idea in Brief
On May 19, 2026, Gartner published its updated worldwide AI spending forecast: $2.59 trillion for the year, a 47 percent increase over 2025. The number made headlines. What did not make headlines was the composition. More than 45 percent of that total — roughly $1.2 trillion — flows to infrastructure: AI-optimized servers, semiconductors, network fabric, and cloud compute. The rest splits among software, services, and models. Almost none of it is earmarked for the organizational work that determines whether any of it pays off.
This would be unremarkable if the infrastructure spending were the hard part. It is not. The hard part — the part that determines whether a $50 million AI program delivers $50 million in value or becomes a $50 million write-off — is integration, governance, change management, and the construction of accountability structures that did not exist before AI arrived. BCG’s research quantifies the split: 10 percent of the effort in a successful AI deployment lies in building or selecting the model, 20 percent in technology and data infrastructure, and 70 percent in transforming business processes and the people who operate them. The spending pattern is almost exactly the inverse of the effort pattern.
The result is not merely waste. It is a structural misallocation that explains much of the enterprise AI failure data that has accumulated over the past two years. McKinsey’s 2025 Global AI Survey found that 88 percent of organizations now use AI in at least one business function, but only 39 percent report any EBIT impact, and only about 6 percent — roughly 109 respondents out of nearly 2,000 — qualify as high performers who attribute more than 5 percent of EBIT to AI. The RAND Corporation’s 2024 analysis found that 80 percent of AI projects fail to deliver their intended business value. MIT’s NANDA initiative found that approximately 95 percent of generative AI pilots never scale to production.
These are not technology failures. They are budget-shape failures — the consequence of an industry that has learned to sell the easy half and an enterprise procurement apparatus that has learned to buy it.
The Easy Half and the Hard Half
The distinction between the easy half and the hard half of AI adoption is not a metaphor. It describes a measurable difference in the nature of the work.
The easy half consists of everything that can be purchased as a product or service with a known price, a known delivery timeline, and a known integration surface. Model licenses. API access. Cloud compute. GPU clusters. Data platform subscriptions. These items appear on vendor invoices, fit into procurement categories that CFOs already understand, and can be benchmarked against competitive quotes. They are, in the language of Oliver Williamson’s transaction cost economics, low-specificity assets: standardized, contestable, and priced by competitive markets.
The hard half consists of everything that cannot be purchased as a product because it is specific to the organization that needs it. Redesigning a workflow so that an AI system’s output connects to the decisions that actually get made. Building a governance structure that assigns accountability when an AI-generated recommendation turns out to be wrong. Retraining a team of professionals whose expertise was built around a process that no longer exists. Constructing the data pipelines, quality controls, and feedback loops that connect a model’s output to the institution’s actual operating environment. These are high-specificity investments — they have no value outside the organization that makes them, they cannot be purchased from a vendor, and their costs are difficult to estimate in advance because they depend on institutional characteristics that vary from firm to firm.
The asymmetry matters because it drives a predictable budget distortion. When a CFO allocates $10 million for “AI transformation,” the portion that can be specified, quoted, and contracted — the easy half — gets funded immediately. The portion that requires internal capacity-building, organizational redesign, and sustained leadership attention — the hard half — gets deferred, under-scoped, or absorbed into existing operational budgets where it competes with everything else.
What the Consulting Firms Actually Found
The evidence for the inversion is now available from three independent sources that converge on the same conclusion.
BCG’s 10-20-70 framework, developed from their work on enterprise AI at scale, holds that 10 percent of the effort in a successful deployment involves algorithms, 20 percent involves technology and data, and 70 percent involves people and processes — adoption coaching, workflow redesign, organizational change management, and the construction of new operating disciplines. BCG’s own research found that companies following this allocation saw measurably higher adoption rates and productivity gains. The framework’s implication is stark: an organization that spends 70 percent of its AI budget on technology and 30 percent on people and process has the ratio backwards.
McKinsey’s 2025 survey isolates the same pattern from the outcome side. The single strongest correlate of enterprise-wide EBIT impact was not model sophistication, data quality, or compute scale. It was fundamental workflow redesign — the deliberate reconstruction of how work gets done, not merely the insertion of an AI tool into an existing process. The roughly 6 percent of respondents McKinsey classified as AI high performers were distinguished not by their technology stack but by their operating model: they deployed AI across more business functions, they ran a full-stack management playbook covering strategy, talent, operating model, technology, data, and adoption, and — critically — half of them intended to use AI to transform their businesses rather than merely to automate existing tasks.
Gartner’s data completes the picture from the market side. Of the $2.59 trillion in worldwide AI spending forecast for 2026, infrastructure accounts for over 45 percent. AI software, services, and models account for most of the rest. Gartner’s own analysts have noted that organizations with successful AI initiatives invest up to four times more in data and analytics foundations than their peers — and that “enterprises have yet to really flex their spending potential” because “CIOs face challenges in proving the value from AI investments and demonstrate tangible business outcomes.” Gartner has placed AI in the “Trough of Disillusionment” for 2026 — a designation that reflects not a technology problem but an organizational-readiness problem masked by technology spending.
The three sources use different frameworks and different data, but they triangulate to the same finding: the cost structure of successful AI adoption is dominated by organizational investment, and the spending pattern of most enterprises is dominated by technology procurement.
The Transaction Cost Explanation
The inversion is not accidental. It has a structural explanation rooted in how institutions make purchasing decisions under uncertainty.
Williamson’s framework distinguishes between transactions that are easy to govern — standardized, frequent, and supported by competitive markets — and transactions that are difficult to govern because they involve asset specificity, uncertainty, and the risk that one party will behave opportunistically once the investment is sunk. The easy half of AI adoption consists almost entirely of the first kind. Model licenses, cloud compute, and API access are standardized products sold in competitive markets with transparent pricing. A CFO can evaluate them using the same procurement logic applied to any other technology purchase.
The hard half consists almost entirely of the second kind. Redesigning a workflow is a one-time investment whose value is specific to the organization. Building a governance structure requires sustained leadership attention from executives whose time is the scarcest resource in the firm. Training a workforce to operate in a new way requires trust, patience, and tolerance for a temporary productivity decline — none of which appear on a purchase order.
The result is that the easy half benefits from all the institutional mechanisms that make large purchases tractable: competitive bidding, vendor accountability, contractual performance guarantees, and the ability to switch suppliers if the first one fails. The hard half benefits from none of them. It requires internal capacity that cannot be outsourced, institutional commitment that cannot be contracted, and sustained attention that cannot be purchased by the hour. When a budget is constrained — and an AI budget is always constrained — the items with clear costs and clear vendors get funded, and the items without clear costs or clear vendors get deferred.
This is not a failure of intelligence. It is a failure of institutional architecture. The procurement systems, budget categories, and approval processes that govern enterprise technology spending were designed for a world in which the technology was the hard part and the organizational adjustment was a minor afterthought. In AI, that relationship is inverted.
The Governance Gap as a Leading Indicator
One of the clearest markers of the inversion is the state of AI governance spending. Gartner estimates that spending on AI governance platforms will reach $492 million in 2026, growing to $1 billion by 2030. Those are significant numbers in isolation. They are trivial relative to the $2.59 trillion in total AI spending they are meant to govern — less than 0.02 percent.
The ratio is revealing. It implies that for every dollar spent on building, deploying, and operating AI systems, less than two-tenths of a cent is spent on the structures that determine whether those systems operate within acceptable risk boundaries. In regulated industries — financial services, healthcare, legal services, pharmaceuticals — the governance requirements are not optional. They are conditions of operation. Yet the spending data suggests that most organizations treat governance as a cost to be minimized rather than a capability to be built.
The parallel to financial regulation is instructive. Douglass North’s institutional economics argues that the rules governing economic activity — formal regulations, informal norms, and the enforcement mechanisms that make them credible — are not costs imposed on productive activity. They are preconditions for productive activity at scale. A financial system without effective governance does not save the cost of regulation; it generates crises whose costs dwarf what governance would have required. The same logic applies to AI adoption at institutional scale. An organization that skips the governance investment does not save money. It accumulates risk that will eventually express itself as a compliance failure, a reputational event, or a quiet erosion of the institutional knowledge that the AI system was supposed to augment.
Gartner’s own research supports this directly: organizations will abandon 60 percent of AI projects lacking AI-ready data through 2026. The Gartner Data & Analytics Summit in March 2026 declared context — the semantic and governance layers that connect AI outputs to business meaning — “the new critical infrastructure.” Four out of five organizations increased their AI investments in 2026, yet only one in five shows measurable return on investment. The gap is not a technology gap. It is a governance and integration gap.
Why the Obvious Fix Is Insufficient
The obvious response is to reallocate: shift budget from infrastructure toward people and process. But the obvious fix encounters a structural obstacle that BCG, McKinsey, and Gartner all name without fully resolving.
The hard half of AI adoption does not have a natural buyer inside the enterprise. Technology procurement has a department. Vendor management has a department. Cloud infrastructure has a department. But “redesign the workflow so that the AI system’s output connects to the decisions that actually get made” does not have a department. It falls between IT and operations, between the CIO and the COO, between the technology budget and the change-management budget. The work is real, the cost is substantial, and the organizational home is undefined.
This is the coordination failure that Herbert Simon described in his theory of bounded rationality: organizations do not optimize across all dimensions simultaneously. They satisfice — they find workable solutions within the decision boundaries they can see. When the AI budget is owned by the technology function, it is optimized for technology outcomes. When the change-management budget is owned by HR or operations, it is optimized for its own priorities. The hard half of AI adoption sits in the gap between these decision boundaries, visible to everyone and owned by no one.
Deloitte’s 2026 State of AI in the Enterprise report names the downstream consequence: pilot fatigue. Organizations that cycle through multiple stalled pilots progressively lose the institutional appetite and cultural momentum needed to complete a production transition. By the third failed pilot, executives stop attending reviews. Champions disengage. The fourth pilot launches into an organization that has already decided, implicitly, that AI does not work here. The accumulated cost is not just the $7 million or more in sunk costs per abandoned initiative. It is the erosion of organizational willingness to try again — a depreciating asset that does not appear on any balance sheet but determines whether the next investment has any chance of succeeding.
The Shape of a Different Budget
The data from the high-performing minority — the 6 percent in McKinsey’s survey, the companies BCG studied that followed the 10-20-70 allocation — suggests what a different budget shape would look like. It would treat organizational readiness, governance infrastructure, and workflow redesign not as overhead to be minimized but as the primary investment that determines whether the technology investment pays off. It would allocate leadership attention — the scarcest and most valuable resource in any institution — to the integration and accountability work that no vendor can do on the organization’s behalf.
This is not an argument against infrastructure spending. The compute, the models, and the platforms are necessary. But they are necessary in the way that a building’s foundation is necessary: essential and insufficient. No one evaluates a hospital by the quality of its concrete.
The question that most CFOs are not yet asking — because the spending categories they inherited were not designed to surface it — is not “How much are we spending on AI?” It is “What is the ratio of our technology procurement to our organizational investment, and does that ratio match the cost structure of successful adoption?” Until that question enters the budget process, the economics of enterprise AI will remain backward: the easy half funded, the hard half deferred, and the failure rates exactly where the data says they should be.
References
Acemoglu, D., & Restrepo, P. (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor. Journal of Economic Perspectives, 33(2), 3–30.
BCG. (2022). Five Rules for Fixing AI and Machine Learning. Boston Consulting Group. (Introduces the 10-20-70 framework.)
BCG. (2025). The Widening AI Value Gap. Boston Consulting Group. (Survey of 1,250 respondents; 60% generate no material value from AI.)
Deloitte. (2026). State of AI in the Enterprise. Deloitte Insights. (Names pilot fatigue and accumulated costs of repeated failed pilots.)
Gartner. (2026, January 15). Gartner Says Worldwide AI Spending Will Total $2.5 Trillion in 2026. Gartner Newsroom.
Gartner. (2026, February 17). Global AI Regulations Fuel Billion-Dollar Market for AI Governance Platforms. Gartner Newsroom. ($492M in 2026, $1B by 2030.)
Gartner. (2026, May 19). Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026. Gartner Newsroom. ($2.59 trillion forecast; Trough of Disillusionment designation.)
McKinsey & Company. (2025, November). The State of AI: Global Survey 2025. McKinsey Global Institute. (88% adoption, 39% EBIT impact, 6% high performers.)
MIT NANDA Initiative. (2025). The GenAI Divide. MIT Initiative on the Digital Economy. (5% of AI pilots achieve rapid revenue acceleration.)
North, D. C. (1990). Institutions, Institutional Change and Economic Performance. Cambridge University Press.
RAND Corporation. (2024). AI Project Failure Rates. RAND Research Brief. (80.3% of AI projects fail to deliver intended business value.)
Simon, H. A. (1947). Administrative Behavior: A Study of Decision-Making Processes in Administrative Organizations. Macmillan.
Williamson, O. E. (1985). The Economic Institutions of Capitalism. Free Press.
Dr. Reza Olfati-Saber is the Founder & Chief Scientist of Wisdom Agent, Inc. His 25+ years of research span the technical foundations of multi-agent AI and the institutional-economics traditions that explain how organizations adopt new technologies. His foundational work is cited more than 49,000 times in academic literature.