The narrative says AI is destroying jobs. Our data says something more precise and more alarming: AI is not eliminating work. It is rerouting it.

We track 2,860 occupations across 39 countries. Every month, we pull employment figures, wage trajectories, vacancy counts, and AI exposure scores from 12 government statistical agencies — BLS, Eurostat, ONS, ABS, Statistics Canada, e-Stat, and six others. We compute a composite risk score for each occupation in each country using a 5-component model weighted by data completeness.

What the Q1 2026 data shows is not a labour market in collapse. It is a labour market in violent reallocation.

The Scoreboard

Start with the headline numbers across our full 39-country dataset:

  • 41% of tracked occupations are rated HIGH or CRITICAL risk — up from 34% six months ago
  • Average global risk score: 52.3 out of 100
  • 96% of occupations now have measurable demand-change signals, giving us near-complete visibility into real-time hiring momentum
  • Confidence in our scores: 95.5% average, meaning almost every risk rating is backed by simultaneous wage, employment, and vacancy data

41% HIGH or CRITICAL is not a rounding error. It represents roughly 1,170 occupation-country combinations where at least two of the five risk components — employment decline, wage compression, falling demand, high AI exposure, or negative momentum — are flashing red.

But here is the number that should get your attention: the occupations rated LOW or STABLE are not distributed evenly. They cluster in specific sectors, specific geographies, and specific skill profiles. The reallocation has a direction.

Where the Risk Is Concentrated

Our country-level data reveals three distinct tiers of exposure.

Tier 1 — Acute risk (average score > 60):

Slovakia (64.9), Latvia (63.4), Estonia (63.1), Bulgaria (63.0), Slovenia (64.1), Luxembourg (63.6). These economies share a common profile: heavy reliance on outsourced professional services — call centres, back-office processing, IT support, financial administration — that were already under price pressure before generative AI made them automatable overnight.

Luxembourg stands out. It has four CRITICAL-rated occupations, the highest count relative to its tracked workforce. Its economy runs on financial services and EU institutional administration — precisely the cognitive-routine tasks where large language models are most immediately cost-competitive.

Tier 2 — Structural transition (average score 50-60):

Sweden (52.4), Denmark (49.5), Iceland (52.0), Croatia (51.8), Germany (48.7), France (47.3). These are mature economies with institutional buffers — collective bargaining, strong unions, retraining programmes — that slow the transmission from AI adoption to employment loss. The risk is real but the velocity is moderated.

The Nordic data is instructive. Denmark and Sweden have among the highest AI adoption rates in Europe, but their risk scores sit below the EU median. The data suggests that institutional infrastructure does not prevent displacement. It buys time.

Tier 3 — Deceptive calm (average score 40-50):

United States (43.1), Canada (44.2), Australia (45.8), United Kingdom (46.1). The anglophone economies show lower average risk scores. This sounds reassuring until you examine the composition.

The US tracks 1,103 occupations — by far the most granular dataset in our system. Not a single one is rated CRITICAL this cycle. But that granularity is precisely the issue: hundreds of highly specific manual and service roles (Tile Installers, Veterinary Technicians, Dental Hygienists) dilute the average. Filter for white-collar professional roles — financial analysts, paralegals, copywriters, data entry clerks — and the US risk profile looks closer to 58.

The Wage Signal

Employment data tells you who has a job. Wage data tells you whether having a job still means what it used to.

Across our dataset, the Eurostat Labour Cost Index reveals a widening gap between headline wage growth and occupation-specific wage compression:

  • Bulgaria: LCI 180.8 (80.8% above 2020 baseline) — but professional services wages growing at half the national rate
  • Romania: LCI 170.5 — similar divergence in cognitive-routine occupations
  • Germany: LCI 117.3 — modest headline growth masking stagnation in administrative and analytical roles
  • Italy: LCI 105.5 — the lowest growth in the EU, but paradoxically less exposed to AI displacement because wages were already too low to make automation cost-competitive

The pattern is consistent across 34 European economies: aggregate wages rise, but the occupations most exposed to AI see their wage growth decelerate. The national statistics look healthy. The occupation-level data tells a different story.

In the US, BLS Occupational Employment and Wage Statistics show the same divergence. Median annual wages for "Computer Programmers" — a role with 0.89 AI exposure on our O*NET-derived scale — have declined in real terms over the past two years. Meanwhile, "Nurse Practitioners" (AI exposure: 0.31) have seen 6.2% annualised wage growth.

The market is pricing AI exposure in real time. It just is not showing up in the statistics most people read.

Ghost Demand

The most counterintuitive signal in our data is what we call "ghost demand."

We track job vacancies across 38 countries using Eurostat JVS, Adzuna, OECD, and national employment agencies. Total vacancy counts have remained roughly stable in most economies. This is often cited as evidence that the labour market is healthy.

It is not.

What we observe in countries where we have both Eurostat vacancy data and Adzuna real-time posting data:

  1. Postings for "AI/ML Engineer", "Data Scientist", and "Automation Specialist" are up 40-60% year-over-year in Germany, France, and the Netherlands
  2. Postings for "Administrative Assistant", "Data Entry Clerk", and "Junior Analyst" are down 15-25% over the same period
  3. The net vacancy count is approximately flat

The aggregate number masks a compositional revolution. The jobs being created are not the jobs being destroyed. The skills required are not the skills being displaced. The wages offered are not comparable.

A displaced administrative assistant in Stuttgart does not become a machine learning engineer in Munich. The vacancy exists. The pathway does not.

The O*NET Map

Our AI exposure scores, derived from the O*NET database's task-level analysis, cover 873 US occupations mapped across all 39 countries. The scores range from 0 (no measurable AI task overlap) to 1 (near-complete task automation potential).

The distribution is not a bell curve. It is bimodal.

Peak 1 — Low exposure (0.1-0.3): Skilled trades, healthcare delivery, protective services, construction, food service. These occupations require physical presence, manual dexterity, or real-time unpredictable human interaction. AI cannot tile a bathroom. AI cannot restrain a combative patient. AI cannot fix a burst pipe at 2 AM.

Peak 2 — High exposure (0.7-0.95): Administrative support, financial analysis, legal research, copywriting, data processing, mid-level management reporting. These occupations consist primarily of information processing, pattern recognition, and document generation — tasks that large language models perform at or above median human quality.

The valley between the peaks — occupations in the 0.4-0.6 exposure range — is narrow and shrinking. These are the transitional roles: some cognitive, some physical, some social. Project managers with hands-on components. Sales roles with technical elements. Teaching positions that combine curriculum design with in-person delivery.

The reallocation is moving workers from Peak 2 toward Peak 1. Not by choice. By elimination.

What Is Actually Growing

Our demand-change data identifies the occupation categories with the strongest positive hiring momentum across multiple countries:

Healthcare delivery: Nurse Practitioners, Physician Assistants, Home Health Aides, Physical Therapists. Every anglophone country and most EU economies show positive demand-change. Ageing demographics and AI's inability to provide physical care create a structural floor.

Skilled construction trades: Electricians, Plumbers, HVAC Technicians, Solar Panel Installers. The green energy transition and chronic underinvestment in trades education have created shortages that AI cannot fill. Germany's Handwerkskammer reports 250,000 unfilled skilled trades positions.

AI-adjacent technical roles: The 40-60% increase in ML/AI postings is real, but the absolute numbers are small relative to the workforce. These roles absorb thousands, not millions.

Care and social work: Mental health counsellors, social workers, eldercare coordinators. Human trust and emotional reasoning remain outside AI's capability envelope.

The common thread: physical presence, manual skill, human trust, or regulatory barriers to automation. The labour market is not dying. It is retreating to the tasks that require a body or a soul.

The Counterintuitive Finding

Here is what our data shows that contradicts the dominant narrative:

Countries with the highest AI adoption rates do not have the highest occupation risk scores.

Sweden has one of Europe's highest AI adoption rates (38% of firms using AI tools, per Eurostat's digital economy survey). Its average risk score is 52.4 — below the EU median. Denmark, another early adopter, sits at 49.5.

Compare with Slovakia (64.9) and Bulgaria (63.0), where AI firm-adoption rates are below 15%. Yet their risk scores are substantially higher.

The explanation is in the composition. High-adoption countries tend to be high-wage economies where AI augments existing workers rather than replacing them outright. The workers are expensive enough that companies invest in keeping them productive. Low-adoption countries have concentrated their professional services in cost-arbitrage roles — the work came there because it was cheap. When AI becomes cheaper still, there is no augmentation story. There is only replacement.

Adoption is not the same as displacement. The countries most at risk are not the ones using AI most aggressively. They are the ones whose economic model depended on humans being the cheapest option for cognitive-routine work.

What We Are Watching Next

Three signals that would escalate our assessment from "structural reallocation" to "systemic risk":

  1. Cross-sector correlation: If risk scores begin rising simultaneously across previously uncorrelated occupation groups — healthcare administration and financial services and creative roles — it means the displacement mechanism has generalised beyond cognitive-routine tasks.

  2. Wage-change acceleration: Not just whether wages decline, but whether the rate of decline increases. A country where wages fell 2% last year and 5% this year is in a qualitatively different situation than one with stable 2% annual decline.

  3. Demand-employment divergence: When vacancy counts rise but employment in the same occupation falls, companies are hiring fewer people to produce more output. This is ghost demand at scale, and it is the leading indicator we trust most.

None of these three signals has fully triggered yet. All three are showing early activation in at least five countries.

The Bottom Line

The Great Reallocation is not a future event. It is the present state of the global labour market.

Our data across 39 countries does not show mass unemployment. It shows mass redirection. Workers are moving — or being pushed — from cognitive-routine roles toward physical, social, and care-intensive work. The transition is painful, uneven, and poorly served by the aggregate statistics that dominate headlines.

The occupations growing fastest pay less than the occupations shrinking fastest. The geographies most exposed are the ones least equipped to manage the transition. The institutional buffers that slow displacement — unions, retraining programmes, employment protection — buy time but do not change the destination.

The data does not tell us where the new equilibrium is. But it tells us the direction of travel. And it tells us we are not there yet.

IsJobSafe tracks occupation-level risk across 39 countries using data from BLS, Eurostat, OECD, ONS, ABS, Statistics Canada, e-Stat, Adzuna, and national employment agencies. Our 5-component risk model is updated monthly. All data is publicly sourced. This is quantitative analysis, not financial or career advice.

Subscribe to receive our research directly. New analysis published monthly.