The 2026 AI labor market is neither collapsing nor booming. It’s splitting into two parallel tracks: a small set of senior technical and “professionalised” roles paying $300,000-plus, and a growing list of routine cognitive roles that AI is hollowing out from the bottom — entry-level first. The headline unemployment rate barely moves. The under-employment rate for workers aged 22 to 25 in AI-exposed occupations is now 19% worse than their peers, up from 13% a year ago.
This piece pulls together what the Bureau of Labor Statistics’ 2024–2034 projections, the Anthropic Economic Index, Stanford’s Brynjolfsson-led entry-level study, PwC’s 2026 AI Jobs Barometer, Revelio Labs’ July 2026 tracker, and the WEF Future of Jobs Report 2025 actually say — and what to do about it. The picture is messier than either the “AI is taking your job” doom or the “AI creates more jobs than it destroys” optimism suggests. If you wanted a clean story, this isn’t it. If you wanted the numbers, here they are.
The 40% number: AI is already in the office, and it’s not coming for your boss
In September 2025, the Anthropic Economics team published the third Anthropic Economic Index report, drawing on millions of anonymised conversations from Claude.ai and first-party enterprise API traffic. The headline finding: 40% of US employees now report using AI at work, up from 20% in 2023 — a doubling in two years. For comparison, the internet took about five years to reach similar adoption rates; electricity took over 30 years to diffuse from urban areas to farm households. The slope of the adoption curve is genuinely without precedent.
What people actually do with AI inside that 40% tells you more than the topline. The September 2025 report (verified 200 from this host) shows coding still dominates at 36% of all Claude.ai conversations. But the second- and third-fastest growing categories are educational tasks (9.3% → 12.4%) and scientific tasks (6.3% → 7.2%). And the share of conversations where users delegate complete tasks to Claude — what Anthropic calls “directive” or automation-oriented use — jumped from 27% to 39% in eight months. People are increasingly handing off full tasks, not iterating.
Geography matters too. The Anthropic AI Usage Index (AUI), measured across 150+ countries, shows Singapore at 4.6x expected Claude usage per capita (driven by a small, highly technical user base), Canada at 2.9x, the United States concentrated in DC and Utah (both around 3.8x). Emerging economies sit at 0.2x to 0.4x. The same pattern that took 30 years to play out for electrification is condensing into three years for AI, and the geographic concentration that produced 19th-century divergence in living standards looks set to repeat.
For the cost of running one of these AI conversations in production, the math is shifting fast — see our breakdown of AI inference cost in 2026 for the per-prompt economics that are driving the substitution math employers are running.
BLS 2024–2034: the official US growth numbers — and the ones nobody quotes
The Bureau of Labor Statistics’ 2024–2034 Employment Projections (verified canonical URL; gov WAF blocks automated HEAD from this host) — released August 2025 — projects computer and mathematical occupations will add 735,000 jobs over the decade, growing 10.1%. That’s the second-fastest growth of any major occupational group, behind only healthcare practitioners and technicians. The detailed-occupation leaders are information security analyst (33% growth, 132,000 openings per year on average), data scientist (36%), software developer (17%), and computer and information research scientist (20%).
Those numbers are real and they matter. But BLS’ same release also projects 13% decline for word processors and related operators, 9% decline for data entry keyers, 16% decline for telemarketers, and continued contraction for cashiers, ticket agents, and travel clerks. The BLS Economics Daily framing is careful: BLS does not explicitly attribute these contractions to AI in the projection methodology, but its companion analysis of “AI impacts in BLS employment projections” does flag that the model assumptions now bake in accelerated automation for routine cognitive work.
The BLS projection is, by design, a ten-year smoother — it can’t capture quarterly shocks. What it does capture is the structural shape of demand over a decade. The shape is: AI-adjacent technical roles grow fast; routine cognitive and clerical roles shrink; everything else is roughly flat. For a complementary data-driven read on what the research consensus actually shows about displacement velocity, the AI jobs research overview from earlier in 2026 lays out the converging signals.
The two-track labor market: PwC’s “professionalised vs democratised” thesis
The most useful framing of the 2026 AI labor market comes from PwC’s 2026 Global AI Jobs Barometer (verified 200), published June 15. They call it the two-track labor market. Track one — “professionalised” jobs — are roles AI is making more expert-intensive: accountants, lawyers, doctors, engineers, designers who use AI to amplify domain judgment rather than replace it. Track two — “democratised” jobs — are roles AI is making easier for non-experts to perform: basic copywriting, simple graphic design, first-draft legal documents, routine data analysis.
The numbers: professionalised jobs are growing twice as fast as democratised ones, and showing 42% higher wage growth since 2021. Companies most exposed to AI are growing headcount faster than least-exposed companies, not slower. The top fifth of most-exposed companies posted 163% productivity growth. This is the opposite of the displacement narrative: when companies commit to AI, they hire more, not less, but they hire differently. They need fewer generalists and more specialists. They need people who can ship production AI systems, not people who can talk about AI in meetings.
For early-career workers, PwC finds the most-exposed junior roles are seven times more likely to demand traditionally senior skills like leadership and strategic thinking, compared to the least-exposed junior roles. “Seniorised” entry-level roles have grown 35% since 2019, even as overall early-career postings flatlined in highly AI-exposed sectors. The traditional career ladder is compressing — junior staff are being asked to operate at a level that took five years to reach in the pre-AI economy. For a deeper look at the workflow-replacement side of this compression, see AI agents are replacing entire workflows.
Where the contraction is actually hitting: Stanford’s entry-level data
The single sharpest piece of evidence on AI’s actual labor effect in 2026 comes from Erik Brynjolfsson, Ruchir Chandar, and Tianyi Chen at Stanford. Their August 2026 update of “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence” — covered in detail by Ars Technica — extends the original 2025 paper with a more recent ADP anonymised payroll sample.
The headline number: workers aged 22 to 25 in the top 40% of AI-exposed occupations are now employed at 19% below the level of their peers in less-exposed occupations. A year ago that gap was 13%. It’s not a fluke — the trend has been consistent across the entire post-ChatGPT period. Drilling in: since 2022, total employment for 22-to-25-year-olds in the most-exposed quartile fell 11%, while their peers in less-exposed jobs grew 10%. Bottom line: this is not a hiring pause or a wage story. It’s a lower employment story, concentrated entirely in the entry level.
The Stanford team uses the Anthropic Economic Index to split occupations by whether Claude usage is “automative” (the model fully replaces a task) or “augmentative” (the model assists a human doing the task). The result is unambiguous: occupations with automative AI use — accountants and auditors, receptionists and information clerks, data entry keyers — show the steepest entry-level declines. Occupations with augmentative AI use — chief executives, registered nurses — show flat or rising employment across all age groups.
The Stanford paper’s interpretation is that AI is primarily substituting for codified knowledge — the kind of formal, standardised, documented knowledge that you can teach through textbooks. Entry-level work is heavy on codified knowledge. Senior work is heavy on tacit knowledge — pattern recognition, judgment, mentorship, stakeholder management. AI substitutes well for the first and complements the second. For a related take on the skills gap this creates, the skills gap piece from earlier in the year covers the upskilling math from a different angle.
For workers in their 30s, 40s, and 50s in the same exposed occupations, the data shows little to no effect. Older workers aren’t being displaced — they’re being augmented. The displacement is concentrated where codified knowledge meets an entry-level salary band. That’s a structurally important fact: the jobs aren’t disappearing from the occupation, they’re disappearing from the career ladder. For workers deciding whether to retrain — and the WEF’s Future of Jobs Report 2025 projects 92 million jobs displaced and 170 million created globally by 2030 — the ladder math matters more than the topline number. See also the broader upskilling data from earlier in 2026 on what retraining actually requires.
The job-postings freefall: Revelio Labs’ July 2026 tracker
Revelio Labs has been publishing its AI Labor Market Tracker monthly since early 2026, and the July 2026 release is the most candid snapshot of where employer demand is actually going. Five headline metrics from the report: job postings in the most AI-exposed occupations have fallen 42% relative to the least-exposed occupations since October 2022, a decline accelerating at -8.6 percentage points year over year. Computer Science undergraduate enrollment has fallen 28% since 2022. 31% of all new professional certifications posted in June 2026 were AI-related; 47% of those AI certifications were in Generative AI / LLM specifically. 5.05 job postings are now required per hire in AI-exposed roles (up 264% year over year). And 93% of the activity change is happening within jobs, not from shifts in the mix of jobs.
That last number is the most important and the least appreciated. When AI changes the content of work, 93% of the change is happening to people who already have the job title, not to new entrants or to workers switching roles. The occupations themselves aren’t disappearing — the tasks inside them are shifting. This matches the Anthropic Economic Index finding that travel agents are being deskilled (complex planning work gives way to routine ticket purchasing) while property managers are being upskilled (bookkeeping tasks give way to contract negotiations).
The matching-market breakdown is the one most directly relevant to job seekers. Revelio finds that the most-exposed occupations now require 5.05 postings per hire, up from roughly 1.4 two years ago — a 264% year-over-year jump. For employers, that means their hiring funnel is broken: jobs go unfilled because candidates are mismatched, candidates get ghosted because recruiters are overwhelmed, and both sides are reporting worse matching experiences. The post-2022 labour market was already tight; AI exposure is making the matching problem worse, not better. The Stanford AI Index 2025 Chapter 4 tracks the same dynamic with LinkedIn data and finds broadly similar patterns: AI job postings growth is decelerating from the 2023 surge, while AI-skilled worker share continues to climb.
One piece of good news in the Revelio data: the supply side is responding. Computer Science enrollment is down 28% since 2022, but AI certifications are exploding. Workers are choosing to upskill rather than re-skill. The challenge is that upskilling takes 6–18 months for most AI-adjacent roles, and the labor market is moving faster than that.
The roles paying $300,000-plus: where the demand has concentrated
If you’re looking for where the wage premium is real and growing, the 2026 data points are unambiguous. AI engineer base salaries are running $227,000–$233,000, with senior total compensation bands of $300,000–$800,000 once you factor equity and signing bonuses. The PwC data shows 56% to 62% wage premiums for AI-skilled workers over non-AI peers in the same occupational category. Anthropic’s January 2026 Economic Index report finds global supply-demand gap of roughly 3.2 to 1 for AI talent. The detailed hiring-and-pay breakdown from earlier in 2026 covers specific company bands and geographic premiums in more detail.
Which specific roles are paying this? Five clusters dominate: (1) Production LLM operations — engineers who can take a model from notebook to production with proper evaluation, monitoring, and cost controls. Hiring managers consistently report this is the hardest role to fill because almost nobody has shipped at scale. (2) AI safety, evaluation, and red-teaming — driven by both EU AI Act compliance work and pre-deployment safety reviews at frontier labs. The Anthropic Economic Index’s January 2026 release introduced new “primitives” around task success rates that map directly to this evaluation work. (3) AI security and adversarial ML — incident-driven demand, especially after the publicised model-extraction and prompt-injection attacks of late 2025. (4) MLOps and inference infrastructure — the boring infrastructure layer that every production AI system depends on, perpetually understaffed. (5) AI-augmented domain experts — accountants, lawyers, doctors, engineers who can ship AI workflows in their own domain. This is the PwC “professionalised” track and it’s where the highest wage growth is concentrated.
What about plain software engineering? BLS still projects 17% growth for software developers through 2034 — but that’s growth in headcount, not growth in starting salaries or junior headcount. The mid-level generalist coding roles that were the engine of 2010s tech hiring are flat or contracting. Coding-bootcamp-to-junior-engineer pipelines that worked in 2018 don’t work in 2026, because coding assistants compress the work that junior engineers used to do. For more on how the junior dev track has specifically shifted, the discussion in mr.technology’s analysis of AI coding assistants and junior developers tracks the same phenomenon from the developer-experience side.
The roles shrinking: not the ones the headlines say
The loudest narrative in 2026 is “prompt engineering is dead.” That narrative is mostly true, but for the wrong reasons. Job postings for “prompt engineer” as a standalone title are down 60% to 79% from the 2023 peak, depending on which labour index you read (Neural Digest, ValueAddVC, Career Skill Guide all converge on that range). What changed isn’t that prompting got less important — it’s that prompting became a baseline competency for AI engineers, product managers, content strategists, and researchers. You can’t ship anything in 2026 without knowing how to write a structured prompt, set up a context window, evaluate outputs, and design a tool-use flow. The skill got absorbed into the job. For the practitioner-side of what changed and what to do next, our “prompt engineering is broken” piece from this week covers the technical shift; the original “prompt engineering is dying” essay at id 20218 covers the broader labour-market framing.
The roles actually shrinking in headcount, per BLS and the McKinsey-style AI exposure index that Forbes synthesised in their August 2026 “58 AI-Exposed Jobs” roundup: cashiers, telemarketers, data entry keyers, word processors and related operators (BLS projects 13% decline), ticket agents and travel clerks, basic copywriters and SEO content writers, customer service representatives in commoditised call centres, and entry-level marketing coordinators. Forbes reports 10,000+ US marketing jobs displaced in 2026 alone, driven by agentic-AI workflows that automate the brief-to-first-draft pipeline.
The Anthropic Economic Index’s January 2026 report adds a useful nuance: it found that Claude shows proficiency on large swaths of certain occupations when measured by task success rate. “Data entry keyers” and “database architects” both have high Claude proficiency, but the implications are opposite. Data entry keyers are mostly codified knowledge that AI is learning to do end-to-end; database architects are mostly tacit judgment that AI is helping to amplify. The BLS projection tells you about headcount. The Anthropic data tells you why. For the broader question of what value AI is actually creating — and where the gains are being captured — the framework in the 2026 AI Value Gap is useful context.
Translator and interpreter roles are also under pressure, though the BLS occupational projection hasn’t caught up to the language-model shift yet. Travel agents have already lost most of their entry-level pipeline. Paralegals doing document review are seeing demand compress. The pattern across all of these is the same: roles where the core work is applying codified knowledge to high-volume standardised inputs are getting automated end-to-end. Roles where the core work involves stakeholder judgment, creative synthesis, or physical-world interaction are largely insulated.
Conclusion: where to position for the next 24 months
The 2026 AI labor market isn’t one market. It’s two, and the gap between them is widening. The aggregate unemployment rate barely budges because the growth in AI-adjacent technical roles roughly offsets the contraction in routine cognitive roles. But the disaggregated picture is stark: entry-level workers in AI-exposed occupations are 19% below their peers, the matching market for those occupations is functionally broken (5+ postings per hire, up 264% year-over-year), and the wage premium for AI-skilled specialists is now 56–62% above non-AI peers.
If you’re entering the workforce, the path that paid off in 2019 — generalist coding, basic copywriting, customer service — is closing fast. The path that’s opening is narrower, more technical, and pays better: production AI engineering, AI safety and evaluation, AI-augmented domain expertise. The window to reposition is open for roughly the next 24 months before the entry-level squeeze hits the next cohort.
If you’re hiring, the data tells you to stop posting generic “AI engineer” roles and start specifying production-experience requirements that match the wage band you’re offering. The matching-market breakdown is the symptom; the underlying cause is that employers are still writing 2019 job descriptions for 2026 work. The candidates who can ship are getting hired. The candidates who can’t are flooding an already-broken funnel.
If you’re investing in education or workforce policy, the data tells you that mid-career retraining is showing higher ROI than undergraduate pipelines. CS enrollment is down 28% since 2022 but AI certifications are up 31% — workers are making the right bet at the individual level even as the systemic answer is still missing. The 92-million-displaced / 170-million-created numbers from WEF are directionally right but temporally wrong: the displacement is happening faster than the creation, and the people being displaced are not the same people being hired for the new roles.
The honest summary: AI is not taking everyone’s jobs, and it’s not creating a new golden age. It’s compressing the entry-level rungs of every ladder it can reach while concentrating the value at the top. The next two years are about whether the labour market can build new rungs fast enough — or whether we end up with a generation of workers stuck between a ladder they can’t climb and a floor that’s been pulled up.
Frequently asked questions
Is AI taking jobs overall or just changing them?
Both, asymmetrically. Across the whole economy, AI’s measured net employment effect is small to flat, per the Revelio Labs July 2026 tracker and BLS 2024–2034 projections. But for workers aged 22–25 in AI-exposed occupations, the gap to peers is now -19% and widening — the contraction is concentrated at the entry level of automative roles. Older workers in those same roles are largely unaffected. The headline unemployment number hides a structurally important under-employment story for early-career workers in routine cognitive jobs.
Will AI replace software engineers?
Senior and specialised software engineering roles are still growing — BLS projects 17% growth for software developers through 2034, and data scientist roles are projected at 36% growth. Mid-level generalist coding roles are flattening. The sharpest contraction is at the entry level: coding assistants compress the junior ramp, and the path from bootcamp to junior engineer that worked in 2018 doesn’t work in 2026. Anthropic’s Economic Index shows coding is still 36% of Claude.ai usage — engineers aren’t disappearing, but the skills that pay are shifting toward system design, AI orchestration, and production LLM operations.
Are prompt engineers still in demand in 2026?
Not as a standalone job title. Postings for “prompt engineer” roles are down 60% to 79% from the 2023 peak, depending on which index you read. The skill — writing structured, context-rich instructions for LLMs — is more in demand than ever, but it has become an expected baseline competency for AI engineers, product managers, and content strategists, not a separate role. The practitioners who do best in 2026 are the ones who absorbed prompting into a broader production skill set rather than specialising in it alone.
What skills are paying the most for AI roles in 2026?
Five skill clusters dominate compensation: (1) production LLM ops / MLOps / evaluation, where the hiring difficulty premium is largest because almost nobody has shipped at scale; (2) AI safety / red-teaming / eval design, driven by EU AI Act compliance and pre-deployment safety reviews; (3) AI security / adversarial ML, where demand is incident-driven; (4) MLOps and inference infrastructure, the perpetually understaffed layer every production AI system depends on; and (5) AI-augmented domain expertise, the PwC “professionalised” track, where accountants, lawyers, doctors, and engineers who can ship AI workflows in their own domain are seeing the largest wage premiums.
Should I still get a CS degree in 2026?
Undergraduate CS enrollment is already down 28% since 2022 per Revelio Labs, but that’s a behavioural response to AI hype, not a labour-market signal. BLS still projects 10.1% growth for computer occupations 2024–34. The degree matters less than what you can ship — AI certifications (especially Generative AI / LLM specialisations) are the fastest-growing credential category, accounting for 31% of all new certifications in June 2026. The strongest signal for hiring managers in 2026 is a portfolio of production AI systems, not the credential itself.