For three years, the question of how many jobs AI is actually eliminating has been answered mostly with anecdotes. A media company cuts its writing staff. A law firm reduces associate hiring by a third. A bank automates a back-office function. Each story is real. None of them, on their own, tell us whether we are watching a transition that will rebalance over time or the start of something more disruptive. Goldman Sachs Research, in an April 24 report by economist Elsie Peng, has now put a number on the question. The number is 16,000 — the count by which AI appears to have reduced average monthly U.S. payroll growth over the past year. It is the first time a major Wall Street research desk has tried to isolate AI’s labor-market signal from the macroeconomic noise around it.

The headline figure has already taken on a life of its own. A handful of secondary outlets have framed it as evidence of a coming labor crisis. Others have dismissed it as a rounding error in a U.S. labor market that adds or sheds hundreds of thousands of jobs every month for unrelated reasons. Both reactions miss what makes Goldman’s analysis interesting. The methodology is the story — specifically, the decision to separate AI’s substitutive effects from its augmentative ones. Once that split is on the table, the conversation about AI and jobs becomes less about a single net number and more about which workers, in which roles, in which geographies, on which timelines.

What did Goldman Sachs actually find?

The methodological breakthrough is the substitution-versus-augmentation split. Earlier studies tended to measure either generic “AI exposure” or “AI displacement” without distinguishing between two very different scenarios: (1) AI replaces a worker outright, eliminating the job, or (2) AI makes a worker more productive, potentially leading to either fewer workers needed for the same output, or more workers as lower per-unit costs increase total demand. Goldman combined an AI displacement score used previously with an IMF-developed AI complementarity index to assign every occupation a position on both axes. The result is a much more granular picture of where AI is destroying jobs and where it is creating them.

Goldman’s framing is unusually careful. Peng explicitly writes that “the aggregate impact of AI on jobs in the past year has likely been smaller than those numbers indicate” because the model does not capture the offsetting effect of data-center construction hiring or the incremental labor demand from AI-related productivity gains. In macro terms, the 16,000-per-month drag should be read as one signal in a much larger picture — meaningful, but not catastrophic. For comparison, the U.S. economy added an average of roughly 150,000 jobs per month over the same period. AI’s drag amounts to roughly one-tenth of average monthly job creation, slowing the pace rather than reversing it.

Which jobs are most at risk?

What is striking about this list is how thoroughly it inverts the historical pattern of automation. Mechanization, computerization, and robotics — the three previous big automation waves — disproportionately hit workers without college degrees in manufacturing, logistics, and manual labor. AI displacement is the first wave to hit college-educated workers in professional services, finance, media, and legal work. The list reads almost like a roster of the standard “white-collar entry-level” pipeline: the roles that aspiring lawyers, analysts, marketers, and operations professionals have traditionally used as on-ramps to senior careers. That those on-ramps are now narrowing is one of the most consequential structural shifts in the report.

Which jobs are AI helping?

The augmentation side of the ledger gets less attention but matters more for understanding where the economy is heading. Goldman’s framework draws on the IMF’s complementarity index, which measures how much an occupation’s tasks require human elements that AI cannot fully replicate — physical presence at a worksite, judgment in ambiguous situations, accountability for ethical decisions, or sustained interpersonal relationships. Customer service representatives and interior designers have similar AI exposure scores in the abstract; what separates them is that the interior designer’s work requires unstructured tasks and physical site visits that AI cannot perform autonomously. So AI complements the designer, where it substitutes for the customer service rep.

This is where Jevons paradox enters the analysis — and where the report becomes genuinely interesting for thinking about the next decade. Jevons paradox, first identified in 1865 in the context of coal consumption, observes that increased efficiency in using a resource often raises total consumption rather than lowering it, because lower per-unit costs expand demand faster than efficiency reduces use per task. Applied to labor: if AI makes an interior designer twice as productive, the firm needs fewer designers per project — but lower per-project costs may attract enough new clients that the firm hires more designers in net. Whether augmentation roles end up adding or shedding jobs in the long run depends entirely on whether demand expansion outpaces productivity gains. Goldman’s data so far suggests it is — augmented roles are a net positive — but it is early.

Why younger workers are bearing the brunt

The most important point about this distribution is that it is not a story about generational adaptability. Gen Z is famously digital-native, comfortable with AI tools, and arguably better-equipped to learn new technology than older cohorts. Their over-representation in displacement has nothing to do with skill and everything to do with role placement. The first three years of any white-collar career are heavy on the exact tasks AI handles best — research synthesis, drafting, formatting, basic analysis, document review. Senior workers have already moved past those tasks into judgment-heavy work that AI cannot replicate. Junior workers haven’t.

Goldman’s economists flag a longer-term concern they call the apprenticeship crisis. Junior roles serve two purposes — they get work done, and they train the next generation of senior workers. If AI eliminates the first function, organizations save money in the short run but inherit a pipeline problem in the long run. Several major U.S. law firms have reduced associate hiring by 25-40% since 2024. Partners report higher margins. The unanswered question: who becomes a partner in 2036 if fewer associates are trained between 2024 and 2030? The same question applies to investment banks, consulting firms, accounting practices, and any profession that historically built senior expertise through years of junior-level pattern-recognition.

The 56% AI wage premium

Three things explain why the premium is widening rather than narrowing. First, demand is outpacing supply. Every organization wants AI-skilled workers; the training pipeline has not caught up. Second, complexity is increasing. Using ChatGPT is trivial. Building reliable AI agents that operate within enterprise governance frameworks is genuinely difficult. The gap between casual users and capable builders is widening, not closing. Third, the stakes are rising. As AI handles more critical business functions, the cost of getting it wrong increases — and organizations pay premium salaries for people who can deploy AI reliably without breaking things. The 56% premium for AI Builders is not a wage signal that will mean-revert quickly; it is a structural mismatch between supply and demand that will likely persist through 2027 at minimum.

What it means for workers, leaders, and the economy

For workers: a practical decision tree

The clearest action for individual workers is to honestly audit two things: what percentage of your current role consists of tasks AI can already do, and which side of the substitution/augmentation split your role sits on. If automation exposure is below 20%, monitor and start building AI fluency, but don’t panic. If it is in the 20-50% range, the right move is to start actively shifting your time toward the non-automatable parts of the work — and toward learning to use AI tools to handle the rest. Above 50%, treat reskilling as urgent. Above 75%, treat it as a career transition, not an upskilling project.

The two paths that actually work, in Goldman’s data: move up the AI value chain — become the person who deploys, manages, and improves AI rather than competing with it; or move into AI-resistant work — roles where physical presence, ethical accountability, complex relationships, or judgment in genuinely ambiguous situations are central to the value delivered. Both paths are viable; which is right for you depends on your starting point. A junior financial analyst is well-placed to move up (becoming an AI-augmented senior analyst). A customer service representative may find moving sideways (into healthcare, skilled trades, or relationship-heavy sales) more achievable than the build-with-AI path.

For tech leaders: workforce planning is now C-suite strategy

The most consequential strategic shift Goldman’s data points to is that AI workforce planning has moved from “an HR concern” to “a board-level strategic question.” Three priorities deserve attention this quarter. First, treat retention of AI-skilled workers as a structural cost. The 56% wage premium will not mean-revert quickly. Organizations that haven’t already adjusted their compensation bands for AI skills are losing their most valuable workers to competitors who have. Second, invest in internal retraining over external hiring. Goldman’s data shows internal retraining costs roughly one-fifth as much per AI-skilled worker, produces workers who already understand the business, and yields 89% two-year retention versus 72% for external hires. Third, plan for the apprenticeship crisis explicitly. If your organization is reducing junior hires to capture AI productivity gains, you are creating a pipeline problem that will manifest in 2031-2035. Structured alternatives — AI-augmented apprenticeships, rotational programs, mentorship-intensive models — preserve the training function even when junior headcount falls.

For the economy and society: real transition, manageable scale

The macro picture is more measured than headline coverage suggests. A 16,000-per-month drag on payroll growth is significant but not catastrophic in a U.S. labor market that adds roughly 150,000 jobs per month on average. Goldman’s own framing — “modest net drag” — is appropriate. The numbers do not, on their own, suggest a wave of mass unemployment. They suggest a transition that is meaningful in aggregate, painful for specific workers and communities, and likely to compound over the next 24-36 months as AI capabilities improve and deployment scales.

Three structural concerns are worth flagging at the policy level. Geographic mismatch: new AI-augmented jobs are not appearing in the same places displaced jobs are disappearing. A laid-off paralegal in Cleveland cannot easily become an AI engineer in San Francisco. Skill mismatch: WEF estimates 59% of the global workforce will need reskilling by 2030, but current training programs reach fewer than 10% of those who need them. Timing mismatch: job destruction happens faster than job creation. AI can eliminate an occupation in 18-24 months. Building the industries, companies, and training infrastructure that create replacement jobs takes 5-10 years. The transition period creates real hardship even if the long-run outcome is net positive — and the burden falls disproportionately on workers and communities least equipped to absorb it.

What happens next

Three things to watch over the next four quarters. First, whether the augmentation effect grows fast enough to offset accelerating substitution. Goldman’s current ratio — roughly 9,000 augmentation gains against 25,500 gross substitution losses for a net 16,000 drag — is the key metric. If augmentation hiring scales as AI tools mature and demand expands, the net drag could shrink. If substitution accelerates faster, it could double. Second, whether the apprenticeship crisis becomes visible in senior-talent metrics. Promotion velocity, time-to-senior-individual-contributor, and the share of leadership roles filled internally vs externally are early indicators. Watch professional services firms first — law, consulting, banking. Third, whether reskilling capacity catches up with reskilling demand. The current gap between workers needing retraining and workers receiving it is the single largest policy question in the labor market.

Final take

Goldman’s report is the most rigorous attempt yet to quantify AI’s labor-market impact in real time. The methodology is sound. The numbers are defensible. The framing is appropriately measured. And the headline figure — 16,000 fewer U.S. jobs per month — has captured public attention in ways that more comprehensive studies have not. That attention is mostly useful, even when it gets translated into more alarming framings than the underlying data supports. The report’s quieter insight — that AI is sorting the labor market into augmentation winners and substitution losers, with younger workers concentrated on the losing side — is the one that should be shaping workforce policy and individual career decisions over the next 12-24 months.

The simplest summary I can offer: This is not the labor crisis some have framed it as. It is also not the painless transition others have hoped for. It is, instead, a slow but accelerating sorting — and the sorting is uneven, structurally biased toward the young and the entry-level, and almost certainly going to compound before it stabilizes. Workers who treat the next two years as a planning horizon for skill investment, employers who treat workforce strategy as a peer to technology strategy, and policymakers who treat reskilling capacity as infrastructure will absorb the transition. Those who don’t, won’t. That is what the Goldman number actually says.