Eighteen Months After
The first wave of AI labor displacement is not showing up where most people expected. It appears in hiring freezes, the quiet closing of the junior-to-senior pipeline, and a professional services job openings rate not seen since 2017. A data-grounded account of what is actually happening.
Everyone expected the layoffs. Mass displacement events — the kind you can point to and say "the AI did this" — were the predicted form of the first wave. That is largely not what has happened. What has happened is quieter, harder to attribute, and in some ways more structurally corrosive than a clean displacement shock would have been.
Eighteen months after the GPT-4 era workplace adoption curve began to register in labor market data, we can describe what the first wave actually looks like. It is a story of closed pipelines, not severed careers. It manifests in who is not being hired, not who is being fired.
What the Data Shows
The pattern these numbers describe is not a recession. Total employment is still growing. Layoffs remain low. The quits rate is flat. By the conventional indicators, the labor market looks "cooling but stable." What the conventional indicators miss is structural: new demand for entry-level knowledge work is disappearing, not through firing, but through not hiring.
How It Is Actually Happening
The mechanisms of the first wave are three, and they are mostly invisible to aggregate statistics.
The labor market is not burning. It is being quietly turned off at the intake valve.
A Snapshot of Specific Sectors
SOFTWARE DEVELOPMENT
This is the most advanced case. The BLS projects computer programmer employment to fall 9.6% over the 2023–2033 decade — the first time a white-collar tech occupation has received a formal negative projection in the modern era. The ADP data confirms the concentration at the junior level. What is less visible: the growth in AI/ML-related roles (BLS projects computer and mathematical occupations overall up 10.1%) is concentrated in the narrow band of workers who build, evaluate, and deploy AI systems — not those who previously used programming as a general professional skill. The field is bifurcating: a small tier of high-demand AI specialists, and a large cohort of general-purpose coders whose market is contracting.
LEGAL SERVICES
Law firms have been the most cautious sector in terms of public acknowledgment, but the signals are clear. Legal research, contract review, and document drafting — the work that historically occupied first and second-year associates — is being compressed. The billable hour model provides structural lag protection: as long as clients pay by the hour, there is an incentive to maintain headcount. But competitive pressure from firms using AI to dramatically reduce research time is eroding this buffer. The junior associate pipeline in major firms is showing early contraction signals that have not yet reached aggregate statistics.
FINANCIAL ANALYSIS AND CONSULTING
The research analyst category — the entry point for finance careers at investment banks, asset managers, and consulting firms — has seen hiring freezes at multiple major institutions. Equity research reports that once required teams of analysts are being produced with smaller headcounts using AI-augmented workflows. The work of a first-year analyst at a bulge-bracket bank — Excel modeling, data compilation, sector research, first-draft memos — maps almost precisely onto the capability profile of current large language models paired with financial data tools.
PROFESSIONAL SERVICES (AGGREGATE)
The JOLTS December 2025 data showing a 225,000 monthly opening decline in professional and business services is the single most important near-term indicator. Professional and business services includes legal services, accounting, management consulting, technical consulting, and business support services. This is the definition of knowledge work. An openings rate at its lowest level since 2017 in this category is not a rounding error. It is a structural signal that employer demand for knowledge workers is under sustained pressure.
The reconciliation is important. The J-Curve model predicted harvest-phase effects in 2027–2029. Humlum's conversion lag (2–4 years) applied to the 2022–2023 enterprise ChatGPT adoption wave puts measurable effects at 2024–2027. The JOLTS data is consistent with early harvest-phase effects that began in 2024–2025. The models are not wrong — they are being confirmed on the earlier end of their predicted ranges.
What the Headlines Miss
The persistent framing in business media is "AI creates as many jobs as it destroys." This is the standard productivity argument — new technology creates new demand that absorbs displaced workers. The historical record is invoked: the automobile replaced horses but created more jobs than it destroyed.
The historical analogy is weaker than it appears. Three specific problems:
The credential reversal. Prior automation waves displaced workers who could absorb into better-paying jobs through education and skill development. AI displacement targets the workers who are already at the top of the skill ladder. The same dynamic that made the automobile-to-auto-mechanic transition work — displaced low-skill workers could learn medium-skill work — does not apply when the displaced workers are already the medium-and-high-skill tier.
The absorption sink problem. Agricultural workers absorbed into manufacturing. Manufacturing workers absorbed (partially) into services. Knowledge workers are the services sector. There is no next sector to absorb into at equivalent earnings and scale. Healthcare and skilled trades are the most commonly proposed candidates, but they cannot absorb millions of displaced analysts, junior lawyers, programmers, and financial professionals without structural mismatches in skills, compensation expectations, and geographic distribution.
The pipeline timing problem. New AI-related jobs are real — BLS projects computer and mathematical occupations up 10.1%. But AI infrastructure jobs (model training, fine-tuning, evaluation, safety) require specific technical skills that take years to develop. A 24-year-old paralegal displaced by contract review AI cannot retrain as a model evaluator in 6 months. The lag between displacement and reabsorption into new AI-adjacent roles is measured in years, not quarters.
What the First Wave Tells Us About the Second
The first wave is a displacement of tasks, manifesting as a compression of roles. Tasks that were previously distributed across junior and senior workers are being consolidated into senior workers augmented by AI. The headcount per unit of output falls. Entry-level positions disappear. Senior positions persist.
The second wave — which our base case model places at 2033–2036 for 50% partial adoption impact — will be qualitatively different. It will not be task compression. It will be role elimination: positions that AI cannot augment but must replace, because the senior workers with the seniority moat have retired or the AI capability has advanced past the moat's protection.
The first wave is real. It is quieter than predicted. It is more concentrated than aggregate statistics suggest. And it is the opening condition for what follows. The workers who are not entering careers today are not a data point. They are the missing generation — the expertise gap that organizations are creating for themselves, 8–10 years forward, in exchange for short-term headcount cost savings.
History is instructive. Manufacturing companies that cleared their workforce through the 2000–2010 shock found, in the decade after, that they had also cleared the institutional knowledge and craft expertise that gave them competitive advantage. The reshoring premium — the extra cost of rebuilding what was dismantled — was measured in years and billions. The knowledge-work equivalent is coming.