The dominant narrative about AI and jobs in the United States has centered on automation eliminating positions. New research from Apollo Global Management's economics team reframes the story: the more immediate effect appears to be wage compression, not mass layoffs. Workers in roles classified as high AI-exposure have seen an estimated 6.7% decline in wage growth since 2023, and the workers absorbing the steepest losses are the ones least equipped to absorb anything.

Roughly 5.8 million U.S. workers, about 3.7% of the labor force, sit in what Apollo classifies as high AI-exposure roles. Most of them are in service occupations. Very few are the tech-adjacent knowledge workers who dominate the AI conversation in policy circles and on earnings calls.

That distinction matters whether you are reading a salary offer or writing one.

What the Apollo research actually found

Published on July 30, the Apollo analysis shifts the AI-and-work debate from headcount to pay growth. The headline figure is a 6.7% decline in wage growth for workers in high-exposure roles post-2023. This is not a nominal pay cut. It is a slower climb, a compression against what workers in comparable roles were seeing before generative AI became a live operational input for employers.

The distributional pattern is where both HR teams and individual workers should focus. The decline was steeper for the lowest earners, particularly those in service occupations. Apollo researchers flagged what they described as having

"potentially profound implications for income inequality and living standards, particularly for lower-wage workers who have the fewest resources to weather the transition."

The framing that treats AI adoption as primarily a headcount story misses what the data is already showing. The question is not only whether roles disappear. It is what those roles pay while they still exist.

Why wage compression is harder to see than layoffs

Layoffs generate a press release. Wage compression generates a quieter offer letter. A candidate who would have received a 5% merit increase in 2022 gets offered 2% in 2026, and the anchor for market rate drifts downward one hiring cycle at a time. No individual notices in the moment. The aggregate shows up eighteen months later in a research paper or a Bureau of Labor Statistics (BLS) Employment Cost Index report.

This is the mechanism worth naming clearly. When productivity gains from AI accrue entirely to the employer without a corresponding wage share, the benchmark that both sides negotiate against begins to slide. Hiring continues. Work continues. The pay curve flattens beneath the surface. Heidi Shierholz, President of the Economic Policy Institute (EPI), has written that productivity gains and wage growth decoupled well before AI became a factor, making the current dynamic a structural acceleration of a longer trend rather than a clean break. That context matters for calibrating how seriously HR teams and policymakers should treat the Apollo numbers.

For employees

If you are in a service occupation or a role that increasingly runs alongside an AI system, three concrete steps are worth taking now.

Benchmark before you negotiate. The market rate a recruiter cites is a moving target, and in AI-exposed roles it may be moving in the wrong direction. Cross-reference at least two independent sources before accepting any anchor. Published sector data from the BLS Occupational Employment and Wage Statistics (OEWS) program, a salary calculator, and current job listings for your exact role are the minimum floor. If the number a recruiter offers is 6% below where the same role sat two years ago, that is a data point worth naming in the conversation, not a floor to accept.

Name the AI adjacency in your negotiation. If your role now includes overseeing, editing, prompting, or quality-checking AI output, that is skilled work with measurable scope. Price it accordingly. Workers who allow AI-adjacent responsibilities to enter their job descriptions without a parallel conversation about compensation are the ones absorbing the compression Apollo measured.

Track your own comparators. Real salary data from peers in your role, anonymized and aggregated, is the most effective counterweight to a recruiter's anchor. The more workers share what they actually earn, the harder it becomes for compression to happen invisibly across an entire occupational category. Under the Fair Labor Standards Act (FLSA), workers have the legal right to discuss their wages with coworkers; no employer policy can lawfully prohibit that conversation.

For HR and compensation leaders

The Apollo finding puts a specific question on the compensation team's desk: is your pay curve for AI-exposed roles reflecting productivity gains, or absorbing them? Answering honestly requires examining at least three years of internal offer data, because the compression Apollo describes is a trend line, not a single event.

Several practical implications follow from the research:

• Benchmark hygiene. WorldatWork's annual Salary Budget Survey and SHRM compensation benchmarking tools are useful inputs, but both carry a 12-to-18-month lag by design. In a compressing market, lagged survey data flatters the offer and understates the gap between what you are paying and what the external market currently commands. Supplement annual surveys with real-time data sources to close that gap.
• Retention risk concentrates at the bottom of the band. The Apollo distributional finding maps directly onto retention risk. Workers with the fewest resources to absorb slower wage growth are also the workers most likely to leave for a marginal raise elsewhere. The cost of one exit at that compensation band, including recruiting, onboarding, and lost productivity, is rarely marginal when calculated fully.
• Financial stress carries a productivity cost. PwC research from April found that employees stressed about money tend to show lower productivity and engagement. If your compensation approach inadvertently deepens that stress in AI-exposed segments, the operating cost lands back on the business, just later and in a different budget line than the one compensation owns.
• Employer brand travels through pay data. Candidates increasingly arrive at interviews with peer salary data pulled from review platforms and compensation calculators. An offer that sits 6% below a plausible benchmark reads as a deliberate position, whether or not it is one. SHRM research has consistently shown that compensation transparency shapes candidate trust before the first interview concludes.

The policy question underneath

The Apollo researchers closed their analysis with a framing that extends beyond any single HR team or hiring cycle. Their conclusion, as reported by HR Dive, is that the critical policy question is not whether AI will reshape the labor market more broadly, but how quickly, and whether workers will have the support they need when it does.

That framing has concrete meaning in a U.S. context. The Department of Labor (DOL) and the NLRB have both signaled interest in how AI adoption intersects with worker rights and bargaining conditions. The Communications Workers of America (CWA), which represents roughly 600,000 workers and has been the most active union in tracking AI's impact on employment terms, has pushed for contract language requiring employers to disclose when AI is used to set productivity targets or compensation benchmarks. Whether or not a given workplace is unionized, those contract provisions point to the questions every worker in a high-exposure role should be asking.

For workers in high-exposure roles, the support that matters most right now is portable skills, current benchmark data, and the confidence to negotiate against a moving anchor rather than accept it. For HR and compensation leaders, the question is whether AI productivity gains will be shared, absorbed quietly, or ignored until the retention data makes the cost undeniable. The 5.8 million workers in the Apollo sample are the leading edge of a broader compensation shift. The decisions employers make now about how to price AI-adjacent work will set the benchmarks that are hardest to reverse.