The $28 Billion Omission: How AI Compresses Wages Without Killing Jobs

The Bureau of Labor Statistics released its quarterly Employment Cost Index on a Tuesday morning in late January. The data showed wages rising at an annualized rate of 3.9%, a figure that mainstream financial media quickly framed as evidence of a resilient labor market. The narrative was neat, comforting, and entirely incomplete. What the headline numbers obscured was a structural shift that no single macro data point can capture: the quiet, relentless compression of individual wage-setting power. Apollo Research, the macro firm run by Torsten Slok, recently quantified this phenomenon at $28 billion annually. That number, when set against a $12 trillion U.S. wage pool, represents a mere 0.23% slice. It sounds negligible. It is not. It is the first measurable signal that AI's impact on labor is no longer a theoretical future risk, but a present-day accounting reality. The code does not lie, but it often omits. What the $28 billion figure omits is the mechanism, the distribution, and the second-order effects that will ripple through the economy for the next decade. This is not a story about robots taking jobs. This is a story about the price of labor being rewritten in real-time. Context: The Paradigm Shift from Displacement to Compression For the past two years, the dominant public narrative around artificial intelligence and employment has been one of substitution. Headlines screamed about the death of the copywriter, the obsolescence of the junior developer, and the coming wave of white-collar unemployment. The data, however, told a different story. U.S. unemployment has hovered between 3.7% and 4.0% for over two years. Initial jobless claims remain at historical lows. The labor market, by every traditional metric, is tight. This apparent contradiction between AI anxiety and labor market stability has puzzled economists and fueled accusations that AI's impact is overhyped. Apollo's research cuts through this paradox with a sharper analytical blade. The impact is not occurring through the quantity of jobs, but through the price of those jobs. This is the distinction between explicit substitution and implicit wage compression. In the explicit model, a worker is replaced by an algorithm. The job disappears, the worker files for unemployment, and the statistic is captured in the jobs report. In the implicit model, the worker remains employed, but their market pricing power erodes. The employer, armed with AI copilots and generative tools that boost individual productivity by 30% to 50%, recalibrates its willingness to pay. The job title stays the same. The responsibilities expand. The salary curve flattens. This is a more insidious, and arguably more consequential, mechanism because it is invisible to traditional labor market metrics. The unemployment rate does not capture it. The JOLTS report does not capture it. The wage growth numbers capture it only in the aggregate, diluted across millions of workers and masked by compositional shifts in the workforce. Core: The Forensic Anatomy of Wage Compression My work as a data scientist involves a similar problem to the one Apollo faced: distinguishing signal from noise in massive datasets. When I analyze on-chain liquidity flows, I am looking for the difference between organic accumulation and wash trading. The same analytical discipline applies to labor market data. The $28 billion figure requires forensic decomposition. Let us build the evidence chain. First, the productivity channel. The deployment of AI tools like GitHub Copilot has been measured to increase coding task completion speed by 55% in controlled studies. In customer service, AI-assisted agents resolve 14% more issues per hour. In content creation, the marginal cost of a first draft has fallen by more than 90%. These are not speculative projections; these are the empirical results from enterprise deployments. When an individual worker's output per hour increases by 30% to 50%, the economic value of that hour changes. The employer now has a choice: pay the same wage for more output, pay a lower wage for the same output, or capture the productivity dividend as profit. The data suggests the third option is dominant. Corporate profit margins are at a two-decade high of approximately 12%, while labor's share of national income has declined from 63% in 2000 to approximately 58% today. The $28 billion annual figure is the first direct measurement of this transfer in the AI context. Second, the distributional asymmetry. The compression is not uniform across the skill spectrum. It is a barbell effect. At the top of the distribution, workers who effectively wield AI tools are capturing a productivity premium. They become the operators of the new machinery, and their wages reflect that scarcity. A senior developer who can orchestrate AI coding agents is more valuable than one who simply writes code. At the bottom of the distribution, workers performing routine cognitive tasks that AI can increasingly handle are facing the strongest downward pressure. The call center operator, the junior paralegal, the entry-level analyst, the first-line copywriter. Their tasks are being augmented or absorbed, and their wage-setting power is eroding. This bifurcation is the hidden story within the $28 billion. It is not a single downward shift; it is a divergence. The Gini coefficient of wage income within occupations is rising, a trend I have observed in the compensation data of tech firms. Third, the entrepreneurship paradox. Apollo's research notes that AI lowers the barrier to entry for new ventures. The cost of building a software product has fallen from millions to hundreds of thousands of dollars. AI-generated code, AI-generated marketing copy, AI-driven customer acquisition. This is real and it is measurable. New business applications in the United States hit record highs in 2023 and 2024. But the same force that lowers the barrier also lowers the moat. When everyone has access to the same AI tools, the differentiation between ventures collapses. The result is a surge in homogeneous, low-quality startups competing on price, driving down the success rate. We are not seeing a renaissance of entrepreneurship; we are seeing a Cambrian explosion of undifferentiated micro-SaaS products, most of which will fail. The $28 billion wage compression figure, therefore, has a dark corollary: a $28 billion transfer from labor income to capital income is financing an entrepreneurial bubble with a high burn rate. Contrarian: The Correlation is Not Causation Before we accept the Apollo framework wholesale, we must apply the skepticism that data demands. The $28 billion figure, as reported, lacks methodological transparency. Was this derived from a model simulation, a survey of employer behavior, or an econometric decomposition of wage data? The coverage is ambiguous. Does it include the gig economy, where the compression effects are likely more extreme? Does it account for the offsetting effects of AI-driven job creation in new categories, such as prompt engineering and AI model tuning? These are not rhetorical questions. The omission of methodology is a red flag. The code does not lie, but it often omits. There is also a conflation risk in attributing all wage stagnation to AI. The period from 2020 to 2024 witnessed the largest inflationary shock in a generation. Real wages fell for 24 consecutive months. The decline in real wages during this period was driven primarily by energy prices, supply chain disruptions, and fiscal stimulus withdrawal, not by AI. To parse the AI-specific contribution from the macroeconomic noise requires a counterfactual analysis that has not been publicly disclosed. Furthermore, the wage compression effect may be partially self-correcting. If AI-driven productivity gains are real and sustained, they should eventually translate into higher corporate earnings, which, through competitive dynamics, could lead to higher wages for the augmented workers who remain. The historical precedent of the IT revolution is instructive. From 1970 to 1995, computerization was associated with wage stagnation for middle-skill workers. But from 1995 to 2005, as the technology matured and complementary investments were made, productivity gains began to flow to wages. We may be in the first phase of a J-curve, where the pain precedes the gain. The contrarian view is not that AI is harmless; it is that the $28 billion figure may represent a transitory disequilibrium rather than a permanent state. The deeper omission, however, is the capital allocation question. The $28 billion extracted from wages flows somewhere. If it flows into reinvestment, R&D, and new hiring in different categories, the net effect could be neutral or positive. If it flows into share buybacks and executive compensation, the effect is a pure transfer of wealth from labor to capital. The data suggests the latter is more prevalent. In 2023, S&P 500 companies spent over $800 billion on buybacks, a record high. This is the on-chain evidence of the economy: the liquidity is flowing to the top. Liquidity flows like water; follow the evaporation. When I analyze a DeFi protocol, I look at where the value accrues. If the protocol's revenue is being drained to the treasury wallet and not distributed to liquidity providers, the protocol is extractive, not generative. The U.S. labor market is exhibiting the same extractive pattern. The productivity gains from AI are being captured by the corporate treasury and distributed to shareholders, not to the workforce that is generating the productivity. This is not a bug in the AI rollout; it is a feature of the current institutional structure. The question is whether this extraction reaches a tipping point where the social contract breaks down. Takeaway: The Signal to Track The $28 billion wage compression figure is not the headline number to watch. It is a lagging indicator, a snapshot of a process that is already underway. The leading indicators are more subtle. I am watching three specific data series over the next 18 months. First, the Employment Cost Index (ECI) for the information technology sector, specifically the wage growth for mid-level individual contributors. If this decelerates below the national average while productivity in the sector continues to rise, the compression thesis is confirmed. Second, the quit rate for white-collar occupations. The quit rate is a measure of worker confidence and alternative opportunities. If the quit rate for professional services falls below pre-pandemic levels while the job opening rate remains stable, it indicates that workers are losing bargaining power. They are staying in jobs they would previously have left because they fear the market has turned against them. Third, and most importantly, I am tracking the ratio of corporate profits to employee compensation in the AI-exposed sectors. This ratio is the cleanest measure of the distributional shift. If it continues to rise, the $28 billion figure will look quaintly small in a year. Code is the oracle; data is the only scripture. The data is telling us that the AI labor revolution is not about job destruction. It is about the redistribution of economic surplus. The question for policymakers, and for workers, is not whether AI will take your job. It is whether the value you create will be allowed to flow back to you, or whether it will continue to be routed to the top. The next ECI release will not answer this question, but it will provide a critical data point. The market is always speaking. We just need to learn to read the compression.