Will artificial intelligence make people unemployed? 20 jobs most at risk

Will Artificial Intelligence Make People Unemployed? 20 Jobs Most at Risk

TL;DR

The labor market is being restructured in real time, and the people absorbing the shock are not the ones writing the press releases. The IMF estimated in January 2024 that nearly 40% of global employment is directly exposed to AI displacement. The retraining infrastructure needed to catch the people losing those jobs does not exist at scale — and the policy architecture was not built for what is already happening.

Key Takeaways

• The IMF found in January 2024 that approximately 40% of global jobs are exposed to AI, with advanced economies disproportionately affected, according to IMF staff research published that month
• The World Economic Forum's Future of Jobs Report 2025 projects 92 million roles displaced between 2025 and 2030, partially offset by 170 million new roles — a net positive that assumes frictionless retraining workers cannot yet count on
• Goldman Sachs estimated in March 2023 that AI could automate tasks equivalent to 300 million full-time jobs globally, concentrated in administrative and white-collar work, according to the bank's internal research published that year
• OECD analysis found approximately 27% of jobs in member countries involve tasks with high automation potential, with clerical, data-processing, and financial support roles at greatest risk
• MIT economist Daron Acemoglu has documented in peer-reviewed research that each robot added per 1,000 workers reduces employment by roughly 0.2 percentage points — automation benefits historically flow to capital owners, not workers
• Klarna disclosed publicly in 2024 that its AI customer service assistant was performing work previously handled by 700 human agents — not a forecast, a disclosed headcount consequence
• Active labor market retraining programs in OECD countries achieve completion rates well below 50%, meaning "new jobs will emerge" remains a projection, not an operational plan

Who Benefits When AI Automates Your Job?

Not you. That is the starting point.

The standard framing for AI and employment goes like this: yes, some jobs will be automated, but new jobs will be created, as happened with the printing press, the industrial revolution, and the personal computer. That framing is not wrong. It is also not useful to the 3.4 million Americans who work as bookkeeping clerks — most of whom are not in a position to wait 15 years for the labor market to rebalance.

The IMF's January 2024 analysis put global exposure at nearly 40%. In advanced economies, where white-collar and knowledge work is more prevalent, the figure climbs higher. The IMF was careful to distinguish between "exposure" (tasks AI can perform) and "displacement" (jobs that disappear). That distinction mostly advantages the people who make the technology, not the people who do the work.

The jobs at greatest risk are not the obvious targets. They are not primarily manufacturing or manual work. They are the administrative, analytical, and communicative roles that sit in the middle of the income distribution. The people who enter data. The people who process insurance claims. The people who write first-draft legal documents, transcribe medical notes, translate technical manuals, and answer customer questions by phone.

These workers are not low-skill. They are mid-skill. And the policy architecture built to support them — retraining subsidies, job placement programs, unemployment insurance — was designed for an economy where displacement happened slowly and regionally. It was not designed for this.

The Data: What the Evidence Actually Shows

The World Economic Forum's Future of Jobs Report 2025 projects 92 million roles displaced by automation between 2025 and 2030, offset by 170 million new roles. That net positive of 78 million jobs is frequently cited. What gets cited less is the mechanism: those new roles require different skills, in different sectors, often in different geographies, acquired over years that displaced workers may not have.

Goldman Sachs put a harder number on immediate exposure in March 2023: tasks equivalent to approximately 300 million full-time jobs could be automated using current AI. That is not 300 million jobs eliminated tomorrow. It is 300 million jobs whose core tasks can now be replicated at a fraction of the cost.

OECD analysis of task-level automation risk found that roughly 27% of jobs in member countries involve predominantly routine or codifiable tasks — the kind AI handles well. Those jobs are concentrated in office administration, financial services, legal support, and data management.

What these numbers share: they are averages. Averages hide the people.

Klarna disclosed in 2024 that its AI assistant was doing work previously handled by 700 human agents. Klarna is one company. Multiply that across financial services and the math becomes uncomfortable quickly.

MIT's Daron Acemoglu, in peer-reviewed research on industrial automation, documented that each robot added per 1,000 workers is associated with a 0.2 percentage point reduction in employment and a 0.42 percentage point decline in wages. The equivalent study for large language models is still being conducted. But the direction of the effect is not unclear.

20 Jobs Most at Risk: A Labor Market Reality Check

The roles below are not speculative selections. Each has documented AI tools already performing its core tasks commercially, today.

Employment figures are drawn from Bureau of Labor Statistics occupational data. Timelines reflect current commercial deployment, not theoretical capability ceilings.

One pattern holds across every category: the roles vanishing first are defined by volume and repetition. The roles surviving longest are defined by judgment, relationships, and accountability — qualities AI can simulate but not be held responsible for.

What This Changes for Journalists, Policymakers, and Labor Economists

For journalists covering this beat: The risk is covering the technology instead of the workers. Every major AI announcement includes a displacement figure in the fine print. Klarna's AI productivity claim is also a headcount claim. Write the second sentence.

Small organizations are now deploying AI agents to handle tasks once done by entire teams — that is the current deployment reality, not a future scenario. If you want a grounded account of what AI agents are actually capable of in small-team contexts today, the gap between marketed capability and documented performance is narrower than most coverage suggests.

For policymakers: The EU AI Act's employment provisions focus primarily on high-risk AI systems and transparency requirements. They do not establish displacement compensation mechanisms. The United States has no equivalent federal framework. The Fed monitors aggregate labor market numbers — unemployment rate, participation rate — which lag real-time sectoral displacement by 12 to 18 months. By the time the data shows up in a Fed report, the workers have already absorbed the shock.

The policy debate is still built on the assumption that displacement is slow, visible, and geographically concentrated — like a factory closure. AI displacement is fast, distributed, and invisible in aggregate statistics until the cohort of affected workers ages out of the workforce entirely.

For labor economists: The standard retraining recommendation needs harder scrutiny. OECD data on active labor market programs shows completion rates well below 50% in most member countries, even before addressing the demand-side question — whether retraining leads to employment at comparable wages, within accessible geography. The jobs workers are being retrained for need to actually exist at the right pay grade. That condition is not reliably met.

How to Track AI's Impact on the Labor Market: A Working Checklist

• Monitor BLS occupational employment data quarterly — specifically clerical, administrative, and financial support categories; volume declines there are the leading indicator
• Track individual company AI deployment disclosures — earnings calls and investor materials contain Klarna-style headcount disclosures that press releases omit
• Watch EU AI Act enforcement activity — the first litigation and enforcement precedents on human oversight requirements will emerge by late 2026 and set the global template
• Follow OECD Employment Outlook editions — the 2025 and 2026 reports will include the first post-LLM cohort retraining data
• Track wage data at the occupation level, not the aggregate — headline unemployment can stay low while specific occupation groups experience real wage compression from AI substitution
• Monitor revenue-per-employee ratios in financial services, legal, and media — the proxy for how much human labor is being removed per dollar of output

Where This Is Heading

The productivity gains will be real. Their distribution will not be automatic.

Every major technology transition of the past two centuries increased aggregate output. None of them automatically distributed the gains to the workers whose tasks were replaced. There is no mechanism in current AI deployment that changes this. The people who own the tools capture the efficiency gain. The people whose work is replaced lose the income. This is not a prediction. It is a description of how every previous comparable transition resolved.

Routine white-collar work is the new factory floor.

The jobs most affected by 20th-century automation were manufacturing roles — predictable, physical, codifiable. The jobs most affected by current AI are the 21st-century equivalent: predictable, cognitive, codifiable. The geography is different — offices, not factories — but the economic logic is identical. And the political response so far has been comparably inadequate.

Aggregate unemployment statistics will mask the impact for years.

Displaced workers who take lower-wage jobs, move to part-time work, or leave the labor force entirely suppress the unemployment rate without appearing as a crisis. Policymakers watching the headline number will be operating on systematically outdated information. Wage data and labor force participation rates by age cohort and occupation are more honest signals — but they receive less attention.

The EU will set the first enforceable labor-protection standards for AI.

The AI Act created the framework. The next EU legislative cycle is expected to address displacement compensation requirements and minimum human oversight mandates in high-volume decision-making — credit scoring, insurance underwriting, legal document processing. The US will follow, probably three to five years later and with considerably weaker provisions.

The retraining industry will grow. Its effectiveness will stay contested.

Bootcamps, upskilling platforms, and government programs are already scaling. Evidence for their effectiveness at scale — specifically whether completion produces durable employment at comparable wages — remains thin and contested. The demand side of the equation is not keeping pace with the supply side. Employers are not hiring newly retrained workers at the volumes and wages the projections assume.

FAQ

Will AI actually eliminate jobs, or will it just change what those jobs require?

Both are happening, and the ratio depends on the role. In data entry, transcription, and routine customer service, elimination is already documented — Klarna's 700-agent disclosure is not unique; it is public. In legal analysis and financial research, AI is changing the task mix without eliminating headcount yet. The rate-limiting factor is not capability; it's organizational inertia and liability. Both will erode.

Didn't previous technology transitions create more jobs than they destroyed?

Yes, historically. The relevant variables are how long the transition takes and who bears the cost during it. The industrial revolution created more jobs than it eliminated over a 50-year arc, but that arc included severe disruption to specific worker cohorts who did not survive it economically. "The long run is positive" is not a welfare policy for the people experiencing the short run.

Does the Fed track AI-driven displacement in its labor market reports?

Not directly. The headline unemployment rate measures people actively seeking work. Workers who take lower-wage employment, move to part-time hours, or exit the labor force entirely disappear from the numerator without registering as displaced. The Fed monitors aggregate conditions. Sector-level wage data and labor force participation rates by age and occupation are more informative for AI impact, but they lag real-time conditions by 12 to 18 months.

Are any sectors completely insulated from this?

Roles requiring licensed physical presence — direct patient care, skilled trades, early childhood education — carry lower exposure. Roles defined by real-time judgment in unstructured physical environments similarly. The common factor is work that cannot be replicated by processing text and data, and where the cost of AI error is borne by a human who would face legal accountability. Neither condition is permanent — it is a function of current capability and current liability frameworks.

What are policymakers actually doing about this?

The EU is furthest ahead. The AI Act establishes transparency and human oversight requirements for high-risk AI systems. It does not yet create displacement compensation mechanisms or sector-specific employment floors. The US has no equivalent federal framework. Several US states have introduced bills requiring disclosure when AI is used in hiring or termination decisions; enforcement is weak. The realistic policy response is running 18 to 24 months behind current deployment pace.

Is retraining workers a viable solution?

Retraining is the recommended solution. It is not yet a proven one at the required scale. OECD data on active labor market programs shows completion rates well below 50% and mixed outcomes on durable employment. The structural problem is demand-side: the jobs that displaced workers are being retrained for need to exist at comparable wages and in accessible locations. That condition is not reliably present in the sectors absorbing the most displacement.

How should I assess whether my specific role is at risk?

Examine the task composition of your role, not the job title. If your daily work centers on processing structured information, applying consistent rules to defined inputs, or producing standardized outputs — written or numerical — the exposure is high. If your role is defined by judgment under genuine uncertainty, accountability to specific individuals, or unscripted physical presence, the exposure is lower. The table in this article maps that distinction for 20 specific occupations. Roles crossing the line from the second category to the first are the ones worth watching most carefully.