The AI Productivity Puzzle: Why Executives See Nothing and Workers See Gains

The numbers arrived with the calm authority of peer-reviewed social science. Eighty-nine percent of business leaders surveyed by Gallup, in research conducted alongside the National Bureau of Economic Research, reported that artificial intelligence had produced no measurable change in their company's labor productivity over the preceding three years. The finding, published via Unusual Whales on 2026-05-22, landed in a policy conversation that has spent years insisting the opposite: that AI is reshaping the economy at speed, that early adopters are pulling ahead, that the productivity revolution is not merely imminent but already underway.

The same survey offered a second data point that cuts sharply against the first. Among US workers in organizations that had implemented AI tools, 65% described the technology's impact on their own productivity as "somewhat" or "extremely" positive. That is not a marginal plurality. It is a two-to-one majority of people who use these systems daily and who say the tools help them work faster, better, or both.

The question is not which number is wrong. Both figures likely reflect genuine responses from genuine respondents. The question is what it means when the people closest to the technology report steady improvement while the people responsible for running the organization see nothing — and what that gap tells us about how productivity is actually being measured, distributed, and captured in the age of AI.

The Measurement Problem

Economists have long known that productivity gains from new technologies arrive unevenly and often with long lags. The steam engine, electrification, and the internet all followed patterns in which adoption preceded measurable output growth by years or decades. The canonical explanation is that complementary investments — new skills, new organizational processes, new management structures — take time to accumulate before the technology's full potential is realized.

AI may be following the same script with unusual fidelity. Eighty-nine percent of leaders reporting no impact does not necessarily mean AI is not working. It may mean that the productivity benefits are accruing to workers rather than to the aggregate output metrics that executives track. If a sales team closes more deals per representative because AI surfaces better leads, the company's total revenue may rise — but if headcount stayed flat or grew, the labor-productivity figure may show little movement. The benefit showed up as margin, not as output per worker.

This is not a minor technical distinction. It goes to the heart of how corporations evaluate technology investment. If the productivity gain manifests as higher margins rather than higher output per worker, it will not appear in the Gallup survey's question about "labor productivity." It will appear in earnings before interest and taxes.

The Distribution Question

There is a second possibility that the data raises and does not resolve: that the benefits of AI are flowing disproportionately to capital rather than labor, and that the 65% of workers reporting positive experiences are describing a real but narrow improvement in their individual working lives — faster drafting, easier research, quicker code review — that is not translating into commensurate gains in organizational output.

The mechanism here is not hypothetical. Studies of automation in manufacturing, logistics, and services have repeatedly found that productivity gains from automation often accrue to owners rather than workers, even when workers report that their individual tasks have become easier or more interesting. The technology increases the productivity of the process; the process remains dependent on human judgment at key nodes; but the financial reward concentrates at the node where capital is deployed.

If this dynamic is playing out in AI-adopting organizations, it would explain both survey results simultaneously. Workers experience genuine improvement in their task-level productivity. Leaders, measuring organizational output per dollar of labor cost, see no corresponding step-change. The technology is delivering private benefits to individual knowledge workers while failing to generate the kind of aggregate productivity uplift that would show up in national statistics or justify the capital expenditure on a balance sheet.

The stakes of this distinction are not abstract. If AI's productivity gains are primarily accruing as margin rather than output, the economic case for aggressive AI deployment weakens considerably. Firms would still invest — competitive dynamics reward any cost reduction — but the argument that AI adoption is a prerequisite for national economic growth would rest on shakier empirical ground.

The Government's Own Problem

The timing of the survey release brought an inadvertent illustration of the broader context. In the same news cycle, a separate report cited via Unusual Whales documented $162 billion in improper payments across 68 federal programs in fiscal year 2024. The figure is a federal government accounting measure — it encompasses overpayments, underpayments, and payments made without proper documentation. It is not a productivity statistic in the strict sense. But it speaks to the same underlying question: whether large, complex organizations can effectively deploy technology to improve performance.

Federal agencies have spent billions on AI-enabled fraud detection, benefit verification, and process automation in recent years. The persistence of a $162 billion improper-payment figure suggests that either those systems are not yet working at scale, or that the problem being addressed is not primarily a technology problem. Improper payments in government programs are overwhelmingly driven by complexity, fragmented eligibility rules, and administrative burden — not by the absence of software. Adding more AI to a system designed by statute to be confusing will not automatically resolve the confusion.

The parallel is precise. Business leaders investing in AI and finding no measurable labor-productivity improvement may be discovering, as federal administrators have found, that the bottleneck is not the technology but the organizational structure surrounding it. AI can make individual tasks faster. It cannot, by itself, redesign the workflow, eliminate the approval layers, or realign the incentives that determine how work actually moves through an institution.

What Remains Uncertain

The Gallup-NBER survey does not explain why the divergence between worker and leader perceptions exists. The data establishes the gap; it does not diagnose it. Possibilities include measurement lag, benefit distribution toward capital rather than labor, complementary organizational change that has not yet been completed, and the mundane possibility that "labor productivity" means different things to a frontline worker and to a chief financial officer.

What is clear is that the dominant public narrative — AI is transforming the economy and early adopters are reaping the rewards — requires qualification. The rewards, where they exist, appear unevenly distributed. They are felt most acutely by the people using the tools. They are hardest to detect in the aggregate metrics that executives and economists use to track national economic performance.

Whether that gap closes, and how, will determine whether the AI investment cycle produces the kind of broad-based productivity growth that would justify its cost in social and economic terms — or whether it produces primarily private returns for firms that are already well-positioned to capture them. The survey released on 2026-05-22 does not answer that question. It does, however, suggest that the question has been insufficiently asked.