The Productivity Paradox: Big Claims Meet Stubborn Numbers
The numbers arrived within hours of each other. First, a survey of business leaders conducted by Gallup in partnership with the National Bureau of Economic Research found that 89% report no measurable impact from AI on their company's labour productivity over the past three years. Then came a federal accountability report showing that $162 billion in improper payments were recorded across 68 programmes in fiscal year 2024 alone. The two datasets concern entirely different domains — one corporate, one governmental — but they share a structure that warrants closer attention.
What connects them is a question that goes beyond the accuracy of any individual statistic: why do large organisations so consistently report vast flows of money and technological transformation while measurable outcomes remain stubbornly flat?
The AI figure deserves scrutiny on its own terms. The industry's marketing cycle has produced a vocabulary of disruption — agents, reasoning models, enterprise deployment at scale — that has become standard in earnings calls, investor decks, and executive communications. Yet the empirical record lags. Gallup's survey, covering a broad cross-section of business leadership, found that the majority had not seen labour productivity shift in any detectable direction over a period long enough to rule out adoption lag as an explanation. NBER's parallel research, published alongside the survey, reached a similar conclusion using a different methodology. The gap between the pace of AI investment and the pace of measurable productivity gain is not a temporary onboarding problem; it has become a durable feature of the enterprise technology landscape.
The government figure is starker. $162 billion is not a rounding error. It represents payments across federal agencies that did not meet statutory, regulatory, or programme specifications — payments made, in many cases, to parties who were not entitled to them. The figure covers 68 separate programmes, suggesting the problem is not concentrated in a single agency or contract type but is distributed across the architecture of federal spending. Improper payments have been a persistent feature of federal financial management for years; the rate has not collapsed despite sustained attention from oversight bodies and congressional committees.
Why do these patterns persist? The incentive structure inside large organisations provides a structural answer. Federal administrators who classify spending as improper rather than as fraud or waste remove the implication of criminal intent while preserving the appearance of programme activity. The metric improves. The underlying process does not. In the corporate context, AI investment is measured by adoption rates, infrastructure spend, and headcount aligned to AI projects — not by output per worker. The gap between process metrics and outcome metrics is not accidental. It reflects the logical structure of internal incentives: it is easier to demonstrate transformation by reporting that money has been spent and systems have been deployed than by waiting for those systems to produce measurable results that may not arrive on a schedule aligned with quarterly reporting cycles.
The structural logic is the same in both cases. When an initiative is too large to fail politically, and too expensive to abandon operationally, the incentive tilts toward protecting the claim of impact rather than demonstrating it. Metrics become management tools rather than measurement tools. The figures are real — $162 billion moved, AI infrastructure deployed at scale — but the translation into outcomes that a detached observer would recognise as transformation remains elusive.
The stakes of this pattern extend beyond any individual programme or technology sector. When institutions routinely report transformation without evidence of it, the informational environment degrades. Decision-makers — whether investors, regulators, or citizens — lose the ability to distinguish genuine performance from reputation management. The credibility of publicly reported metrics erodes across the board, making it harder to allocate capital efficiently, to design accountability systems that work, or to hold power to account. In a wartime economy like Ukraine's, where the state's credibility is measured not only in exchange rates and infrastructure but in the capacity to protect civilian life from glide-bomb strikes, the cost of this degradation is not abstract. It has a direct human dimension.
What the available data does not explain is why the gap between institutional claims and institutional results has widened rather than narrowed as measurement infrastructure has improved. Both the AI productivity gap and the federal improper payment rate are documented phenomena — studied, reported, and discussed. Neither has produced the structural reform the evidence seems to demand. The most parsimonious reading is that the gap is not a bug in the system. It is the system.
The survey data and the accountability report represent two of the more precise datasets published this week. They tell a coherent story: transformation is being reported at a scale that the underlying evidence does not support, and the incentives maintaining that gap are structural rather than incidental. Whether that story prompts institutional correction or simply gets filed alongside the rest of the quarter's data will say more about the health of accountability mechanisms than either dataset alone.