The Two Numbers Nobody Wants to Explain Away

The AI revolution was supposed to be everywhere at once. It is nowhere to be found in most places that matter.

That is the blunt conclusion buried inside new survey data released this week: 89 percent of business leaders report that AI has had no measurable impact on their company's labor productivity over the past three years. The figure, drawn from a combined Gallup and National Bureau of Economic Research survey and reported by Unusual Whales, is not a minor data point. It is an indictment of an entire discourse — one that spent years insisting the productivity gains were imminent, then arriving, then inevitable.

Meanwhile, the United States government reported on the same days that it had recorded $162 billion in improper payments across 68 federal programs in fiscal year 2024 alone. That figure — also carried by Unusual Whales, citing the governmentwide accountability report — represents one of the largest single-year waste totals on record.

Two numbers. One lesson. The institutions that promise transformation most loudly are, by some structural logic, the least likely to deliver it.

The AI Productivity Mirage

The Gallup-NBER finding needs to be stated plainly: after three years of generative AI deployment at scale, nearly nine in ten senior leaders at American companies say the technology has not moved the needle on output per worker. This is not a marginal miss. It is a near-total failure of stated purpose.

The standard explanations offered — implementation friction, workforce reskilling gaps, integration complexity — are themselves revealing. They describe a technology that works in isolation but falters when placed inside an organization. That is not a technology problem. That is an institutional problem wearing a technology costume.

The incentive structures inside large organizations actively resist productivity gains. A middle manager whose headcount budget is her measure of power has no interest in a tool that renders her team redundant. A procurement department whose influence scales with vendor count will find ways to slow adoption. The productivity gains exist in the lab; they dissolve on contact with organizational immune systems.

This publication has noted before that the gap between a technology's capability and its deployed impact is a function of governance, not engineering. The survey data on AI this week is the most recent confirmation of that pattern.

The $162 Billion Leak

The government improper payments figure is more visceral but less discussed. $162 billion — across 68 federal programs in a single fiscal year — does not represent waste in the abstract. It represents a structural incapacity that has persisted, and grown, despite decades of oversight reform.

The programs affected span Medicaid, unemployment insurance, federal student loans, and Medicare. In each case, the root causes are different: ineligible recipients, duplicate payments, documentation failures. But the common thread is administrative entropy — systems designed to process volume, not to verify truth.

Congress appropriates. Agencies spend. Oversight exists, but in fragmented form, attached to programs that were never designed to be accountable to it. The incentive is to distribute, not to verify. Verification costs money and slows disbursement; politicians who slow disbursements get attacked for bureaucratic cruelty. The result is a system that optimizes for throughput over accuracy, year after year.

The irony is that AI is being pitched as a solution to exactly this category of failure — fraud detection, pattern recognition, anomaly flagging. If the productivity data holds, the technology intended to fix the leak will not work either.

Why Institutions Cannot Self-Correct

The structural commonality between the AI failure and the government waste figure is the absence of genuine accountability pressure. In the corporate sector, AI deployment is measured in announcements and pilot programs, not in quarterly productivity delta. A CEO can claim to be leading an AI transformation without a single verified output gain. The board lacks the tools to distinguish posture from performance.

In government, the accountability structure is nominally present but practically toothless. Inspectors general produce reports that are read by nobody with the power to act. GAO recommendations accumulate in archives. The $162 billion figure appears; congressional hearings are held; witnesses testify; nothing changes at the program level.

Both failures share a deeper root: the institutions that control the deployment of AI and the allocation of government money are the same institutions that benefit from the status quo. The technology promises productivity, but productivity is threatening to the people whose influence is measured in headcount. The programs promise efficient delivery, but efficiency is threatening to the constituencies whose access depends on administrative friction remaining intact.

The Reckoning That Is Coming

At some point — the data suggests that point is now — the gap between institutional promise and institutional delivery becomes politically untenable. The public knows AI is everywhere and nowhere useful. Taxpayers know government spending produces results that cannot survive basic auditing. The credibility of both the technology industry and the federal government is eroding on the same timeline.

What replaces that credibility is the open question. One path is structural reform: tying AI deployment bonuses to verified productivity metrics, centralizing improper payment detection with real-time program authority, breaking the organizational immunity that protects managers who underdeliver.

The other path is inertia dressed as reform — more announcements, more task forces, more dashboards that track activity instead of outcomes.

The numbers this week are a vote for neither party. They are a vote against the assumption that scale and investment automatically produce results. That assumption has sustained both the AI hype cycle and the government spending machine for years. The evidence, such as it is, suggests the assumption deserves to be retired.

Monexus covered the AI productivity survey and the government improper payments report as two distinct data points. This article argues they are, in fact, one story.