The AI Productivity Paradox: Why the Technology Touted to Transform Work Is Leaving Most Workers Behind
Three years into the most concentrated corporate adoption campaign in technology history, almost nine in ten business leaders say artificial intelligence has made no measurable difference to their company's labor productivity. That is the central finding of a Gallup survey conducted in partnership with the National Bureau of Economic Research, released on 22 May 2026 through the research platform Unusual Whales. The number is not a rounding error or a marginal result. It is a near-unanimous verdict from the people responsible for deploying the technology.
The survey's publication on the same day President Donald Trump announced the suspension of an executive order governing artificial intelligence oversight added a public-policy dimension to what corporate boardrooms have privately been wrestling with. When the leader of the free-market economy pauses governance of the technology most frequently cited as the defining commercial force of the decade, and the companies most aggressively adopting it are simultaneously reporting that the technology is not working, the coincidence demands scrutiny.
The numbers do not exist in isolation. A report from the Government Accountability Office, also surfaced through Unusual Whales on 21 May, identified $162 billion in improper payments distributed across 68 federal programs during fiscal year 2024. Federal agencies, like private firms, have been under pressure to integrate automation and algorithmic systems into disbursement workflows. The scale of misdirected public money raises a parallel question: if automated systems were the answer, who was watching the machines?
This publication finds that the productivity gap, the oversight vacuum, and the federal payment failures share a common structural thread. Each represents a different failure mode of the same underlying problem: the gap between the institutional enthusiasm with which AI is adopted and the rigor applied to measuring whether it actually works.
The Productivity Audit Nobody Wanted
The Gallup-NBER survey asked a specific question. Across a sample of business leaders — executives, senior managers, and principals responsible for technology procurement and workforce planning — 89 percent reported no net change in labor productivity attributable to artificial intelligence over the preceding 36 months. Nine percent reported a positive impact. The remaining respondents were uncertain or declined to answer.
The finding complicates a narrative that has dominated technology-sector public relations and investor communications since the release of large language model tools in late 2022. The promise embedded in that narrative is straightforward: AI will augment human workers, eliminate repetitive cognitive tasks, and unlock efficiency gains that translate to top and bottom-line growth. The survey data suggests that promise has not, on average, been redeemed in the timeframe corporate planning cycles require.
There are several explanations available, each grounded in plausible mechanism rather than speculation. The first is implementation lag: companies acquired access to the tools but have not yet reorganized workflows to extract their value. This is a standard diffusion-of-innovations argument and would predict that the productivity effect will arrive, just later than marketed. The second explanation is measurement failure: conventional productivity metrics, designed for industrial-era output, may not capture the value AI generates in advisory, analytical, or creative contexts where output is harder to quantify. A consultant using AI to cut research time by half may show no productivity gain on the firm's billable-hours metric. The third explanation is the most uncomfortable one: the technology, in its current enterprise-grade form, may simply not be as capable as its advocates claim for the specific tasks companies are asking it to perform.
What the survey does not do is adjudicate between these explanations. Its value is in establishing that the question is now empirically unavoidable. Three years of corporate deployment. Eightynine percent reporting no productivity effect. The burden of proof has shifted.
Government Money, Automated Mistakes
Fiscal year 2024 produced $162 billion in improper federal payments across 68 programs. The figure comes from federal accounting records as compiled and reported through Unusual Whales. It represents payments made in error — to ineligible recipients, for unfulfilled services, in incorrect amounts, or without adequate documentation. The programs span Medicaid, unemployment insurance, student aid, Medicare, and a range of smaller grant mechanisms.
Improper payment rates are not new. The Government Accountability Office has tracked them for decades, and every administration inherits a legacy of programs that distribute money faster than they verify it. What is new, in the current cycle, is the pressure applied — by oversight bodies, by congressional committees, and by the administration's own deregulatory posture — to deploy artificial intelligence in exactly these disbursement workflows.
The logic is coherent on its face: a system that routes applications through automated eligibility verification, cross-references applicant data against multiple federal databases, and flags anomalies for human review should, in theory, reduce improper payment rates. The experience of the past three years suggests the theory has run into the particular resistance of administrative reality.
Federal benefits programs serve populations whose circumstances — income, household composition, medical status, employment — change frequently. Automated systems trained on historical patterns can discriminate against applicants whose situations do not fit prior distributions. They can perpetuate the assumptions embedded in their training data. They can also fail catastrophically when the underlying program rules change and the system has not been updated. Each of these failure modes has been documented in recent years across Medicaid redetermination processes, unemployment insurance systems during high-claim periods, and student loan repayment programs.
The $162 billion figure is not, in isolation, proof that automation caused the losses. Some portion of improper payments pre-dates AI adoption in these programs. But the scale of the number, emerging alongside a political moment in which the executive branch is simultaneously expanding AI deployment mandates and rolling back oversight mechanisms, makes the structural risk more acute, not less.
The Oversight Vacuum and Its Consequences
On 21 May 2026, President Trump announced the suspension of an executive order governing artificial intelligence oversight, telling reporters he did not like certain aspects of the directive. The order, whose specific provisions were not enumerated in the available reporting, had apparently established some framework for federal review of AI systems — likely touching on procurement standards, risk classification, or interagency coordination.
Its suspension leaves a gap. Federal agencies have been moving, under various legislative and administrative mandates, to integrate AI into decision-making processes that range from benefits administration to contract award to environmental review. Absent a coordinating oversight framework, each agency is left to develop its own standards — creating the conditions for uneven quality control and for the kind of system-specific failures that produce aggregate improper payment tallies.
The gap is not merely procedural. It is philosophical. The question of who bears risk when an AI system produces a wrong outcome — the agency that deployed it, the vendor that built it, the contractor that implemented it, or the civil servant who approved the procurement — has never been clearly resolved in American administrative law. An oversight executive order, whatever its specific provisions, represented at minimum an attempt to begin that resolution. Its suspension signals that the resolution is being deferred.
Deferral is not a neutral act. It advantages actors with the resources to develop internal AI governance capacity — large technology firms, well-staffed federal agencies, and contractors with compliance infrastructure — and disadvantages those who lack it. In the context of the $162 billion improper payment figure, deferral means that programs serving low-income and high-vulnerability populations will continue to rely on AI systems without any federal framework defining what adequate human review looks like, what error rates are acceptable, or what recourse beneficiaries have when automated decisions go wrong.
What the Evidence Actually Shows
The combined picture is not one of technology that does not work. Language models can generate fluent text, synthesize large document collections, and produce functional code. The failure being documented is narrower and more specific: the technology is not reliably producing measurable productivity gains in the business contexts where it is most aggressively deployed, and government programs deploying it at scale are producing significant financial errors without a governing framework to manage those errors.
This is a different claim than AI being a failure. It is a claim about the distance between what AI does in demonstration conditions and what it does embedded in organizational workflows, against existing performance benchmarks, measured with existing metrics. That distance is real. The Gallup-NBER survey makes it empirical rather than anecdotal.
The policy implications are equally specific. Effective AI governance does not require the technology to be banned, regulated into irrelevance, or treated as a strategic threat. It requires the institutional discipline to measure whether the technology is doing what it is claimed to do — and to act on the measurement results, including by stopping deployments that are not delivering value or are producing harms that outweigh it.
The pause of the AI oversight executive order moves in the opposite direction. It removes a coordination mechanism at the precise moment that the evidence of implementation failure is becoming harder to ignore. Whether the suspension is temporary or signals a more fundamental retrenchment from federal AI governance is not yet clear from the available reporting. What is clear is that the productivity survey, the improper payment tally, and the oversight pause are not separate stories. They are three facets of the same unresolved question: what does accountability look like when the machines are making the calls?
This publication covered the AI productivity story through the Gallup-NBER survey as the primary empirical anchor, using the Government Accountability Office improper payments report as corroborating evidence of automation governance failure, and the executive order suspension as the policy dimension that makes the stakes concrete rather than abstract. The framing in dominant wire coverage tends to treat these as separate regulatory and corporate-interest topics. Monexus has connected them structurally.