The AI Spending Paradox: Why Billions Are Buying So Little Productivity

Something is badly wrong with the mathematics of artificial intelligence. Companies have poured hundreds of billions of dollars into AI systems, signed multi-year infrastructure contracts, restructured entire workforces around machine-learning tools — and yet, according to a Gallup survey conducted for the National Bureau of Economic Research, 89% of business leaders report that AI has produced no measurable improvement in labor productivity within their organisations over the past three years.

That is not a rounding error. It is not a lag between investment and payoff. It is a structural data point that demands an accounting.

Meanwhile, the US federal government reported $162 billion in improper payments across 68 programs in fiscal year 2024 alone. The agencies spending that money are doing so, in many cases, inside the same AI deployment wave that the private sector is finding so unproductive — and with far less institutional scrutiny than publicly traded companies face.

The result is a compounding accountability vacuum: capital being allocated at a scale never seen before, against outcomes that cannot be demonstrated, by institutions that have been deliberately weakened against the possibility of oversight.

The Productivity Illusion

The AI industry's growth story rests on a simple premise: invest in the infrastructure, the models will improve, the productivity gains will follow. For frontier model developers — OpenAI, Anthropic, Google DeepMind — that logic has produced enormous valuations and a near-monopoly on the public imagination. For the corporations deploying their products, the evidence is considerably cooler.

The Gallup data, released by Unusual Whales on 22 May 2026, records that nine in ten surveyed executives at companies with active AI programs found no statistically significant improvement in output-per-worker over a period long enough to absorb learning curves, integration friction, and early-stage adjustment. The survey covered organisations across sectors, company sizes, and levels of AI deployment maturity.

What is less surprising is the direction of travel. AI adoption in corporate environments has followed the pattern of most general-purpose technologies: initial over-investment relative to demonstrated use-cases, a period of disillusionment, then slower but genuine integration into specific workflows where it is competitive. The companies that spent most aggressively in 2022 and 2023 are now finding that software licenses and GPU infrastructure do not automatically produce worker output gains. The models are genuinely better than they were. The organisational change required to capture that improvement has proven slower and more expensive than the sales pitch suggested.

The Federal Liability

The government waste figure adds a different dimension to the problem. $162 billion in improper payments in a single fiscal year — money paid out incorrectly, fraudulently, or without adequate documentation — is not new. Federal agencies have reported this kind of figure for years. What is new is the combination of that spending with aggressive AI deployment inside agencies that lack the audit infrastructure to evaluate whether the systems are working.

The accountability problem is not abstract. AI procurement in federal agencies involves large contracts with a small number of vendors, limited competitive bidding, and oversight mechanisms — inspectors general, GAO reviews, congressional appropriations committees — that have been stretched thin across multiple policy priorities simultaneously. If private companies cannot demonstrate productivity gains from AI, the federal government is not positioned to do better. It is positioned to spend the money without the scrutiny.

The political context matters here: a period of aggressive federal cost-cutting has reduced the institutional capacity of oversight bodies. When the question is whether an AI system deployed inside the Social Security Administration or the Veterans Affairs data infrastructure is producing the outcomes its vendor promised, the capacity to ask that question and get a credible answer has narrowed. That is a different kind of risk than the productivity problem in the private sector — it is not about returns on investment, it is about the basic functioning of government services.

Why the Story Is Being Told Wrong

Media coverage of AI investment has settled into a pattern of aggregating capital commitments — $80 billion here, $100 billion there — and treating the total as a proxy for progress. That framing is appealing because it is simple and because the companies doing the spending have strong incentives to reinforce it. A large infrastructure commitment signals confidence and attracts further capital. It does not, however, say anything about whether the systems being built are producing value.

The Gallup data cuts across that narrative because it comes from the people making the deployment decisions, not from industry advocates or academic projections. 89% of leaders saying the technology is not working in their organisation is not a prediction about the future. It is a verdict on the present.

The structural issue is that productivity measurement has not kept pace with what AI actually does. Output-per-worker is a poor metric for a system that improves the quality of decisions, accelerates the rate of iteration, or reduces the cost of new products rather than producing more of the same ones. Companies that measure the wrong things will conclude AI does not work. Companies that find the right measurement frameworks — and have the patience to use them — may eventually find that it does. The problem is that neither the market nor the political system currently rewards that kind of disciplined inquiry.

What Comes Next

If the Gallup figure is real and persistent — if 89% of leaders continue to report no productivity impact into 2027 and 2028 — the investment cycle will eventually break. Companies will stop signing infrastructure contracts, the valuations of the hyperscalers will compress, and the political narrative around AI will shift from inevitability to reckoning. That reckoning would be appropriate. It would also be incomplete without a parallel examination of what is happening inside government.

The federal government does not face a market test. It faces a political test, and the political test has been structured to be lenient. $162 billion in improper payments is not a scandal in the current environment — it is a background condition. AI systems deployed without functional oversight are not a risk that anyone in the executive branch is being held accountable for managing. That changes the calculus: a productivity collapse in the private sector would produce a market correction. A productivity collapse in federal AI deployment would produce slower, less visible, and harder-to-rectify failures in the services that citizens depend on.

The data points in one direction: enormous sums are being spent, against outcomes that cannot be demonstrated, inside institutions whose capacity to ask hard questions has been reduced. The appropriate response — for investors, for regulators, for boards, for anyone asked to fund the next phase of AI infrastructure — is to stop treating the spending as evidence of progress and start treating the absence of productivity data as the warning it is.

This publication covered the AI productivity gap from the angle of institutional accountability rather than the technology investment narrative dominant in business wire reporting.