The AI Productivity Mirage Is a Governance Problem, Not a Measurement Problem

Something is not adding up. Companies have poured hundreds of billions into artificial intelligence, executives have declared AI transformation their top strategic priority, and yet, according to a Gallup and NBER survey reported on 21 May 2026, 89 percent of business leaders say AI has made no measurable difference to their company's labor productivity over the past three years. The obvious framing is that we're measuring the wrong things — that productivity metrics are too blunt, that the gains are intangible, that the tools are ahead of the processes. That framing is comfortable. It lets everyone off the hook. It is almost certainly wrong.

The more uncomfortable interpretation is structural: the AI rollout is not primarily a productivity project. It is a capital allocation exercise, a hedge against competitive obsolescence, and a performance for markets and boards. If it were a productivity project, we would be asking harder questions about where the returns are, why implementation is failing, and who is accountable for the gap between promise and outcome. Instead we are told to wait — next year, the gains will materialise. The measurement frameworks will mature. The integration will deepen. This has been the line since 2023.

The savings and the silence

The survey finding is not an outlier. It sits alongside a pattern that researchers and watchdogs have documented across government and industry: large-scale technology investments routinely produce impressive rollout narratives and thin evidence of actual efficiency gains. A report from the Government Accountability Office identified $162 billion in improper payments across 68 federal programs in fiscal year 2024 alone — a figure that reflects not merely fraud but the operational failure of systems that were supposed to be modernised. The technology was supposed to fix this. By the metrics being used internally, it may have. By any independent measure of actual performance, it has not.

The parallel is instructive. When government agencies misspend tens of billions, the post-mortem usually points to incentive misalignment, inadequate oversight, and a procurement culture that rewards vendors for delivering contracts rather than outcomes. The private sector is not immune to these dynamics — it is, in many respects, the template. Tech vendors sell the transformation; consulting firms charge to implement it; boards approve the budget; the metrics by which success is judged are set by the same institutions that sold the project. It is a closed loop, and it produces exactly the results we are seeing: enormous spending, confident pronouncements, and silence where the productivity numbers should be.

Why the AI case is different in kind, not just scale

The standard defence of AI adoption is that it is still early. Every general-purpose technology — electricity, the internet, electrification of manufacturing — took a decade or more before aggregate productivity gains appeared in the data. This argument has a surface logic. But it confuses two different kinds of delay. In prior technology cycles, the delay was in adoption and implementation: the productivity gains existed as soon as the machines were deployed and the workflow redesigned, but only a fraction of firms had done that work. The gains were real; the diffusion was slow.

AI is different in a specific and important way: the technology requires no physical workflow redesign to be deployed. A language model does not need a factory floor reconfiguration. A chatbot can be plugged into a customer service operation in days. The bottleneck is not technical integration — it is organisational, cultural, and political. The productivity gains are not waiting on a diffusion curve. They are latent in tools that are already inside companies and producing, apparently, very little. That is not an adoption lag. That is a value extraction failure.

The Gallup and NBER data captures this directly. The 89 percent figure is not a complaint about early days — it is the report of executives who have been living with the tools and finding them wanting. They are not saying the technology is too new. They are saying it is not working. The distinction matters because it directs the question away from patience and toward accountability.

What the gap costs and who pays it

The financial scale is substantial. Global AI spending is running at an estimated $600 billion annually and climbing. If 89 percent of that investment is producing no measurable productivity effect, the misallocation is measured in the hundreds of billions per year. That is not a rounding error. It is a structural distortion of capital markets, comparable in scale to the government waste figures — and it has a similar distribution of accountability. The investors and shareholders absorb the loss indirectly, through compressed margins and redirected R&D budgets. The tech vendors absorb nothing. The executives who championed the investments face no personal consequence as long as the board is satisfied with the narrative. The workers whose jobs are being restructured around AI tools that do not make them more productive absorb the friction, the precarity, and the uncertainty.

There is a policy angle here too. When private capital misallocates at this scale, it redirects talent, political attention, and regulatory capacity away from problems that do have demonstrable solutions. Infrastructure, healthcare, education, and industrial policy are all competing for the same ecosystem of expertise and investment appetite. An AI arms race that produces no returns crowds out that competition in ways that are not easily reversed. The opportunity cost is invisible in any single quarter; it is stark across a decade.

The right question

The conversation about AI productivity has been dominated by a false choice: either the technology works and we need better metrics to see it, or the metrics are wrong and the technology works. The Gallup and NBER survey suggests a third possibility — that the technology is being adopted for reasons that have little to do with productivity, and that those reasons are doing exactly what they were designed to do. Companies are hedging, performing, and repositioning. They are not producing more with less. The fact that they are not willing to say so explicitly is itself a data point about how this technology has been sold and how the politics of corporate adoption work.

The honest question is not whether AI has a productivity problem. The question is whether the institutions deploying it have an accountability problem. Until that question is asked with the same urgency that the technology receives, the gap between investment and outcome will remain a feature of the system, not a bug to be corrected.

This publication's coverage of AI infrastructure spending and productivity data reflects a pattern Monexus has tracked across both private and public sectors — the metrics companies use to declare success often diverge sharply from those that would measure actual output or value. The data cited here is drawn from NBER's published survey instruments and the GAO's fiscal year 2024 improper payments report.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

• https://t.me/TSN_ua/25417