Most discussion of AI and jobs runs on a single number: how much of an occupation's task content a model can perform. That number identifies where technology reaches. It does not say what happens next. An exposed occupation may shed workers, or it may keep them while the work itself is rebuilt around the technology. Those are different outcomes, they call for different policy, and exposure alone cannot tell them apart.
Applying the Richmond (2026) framework to Türkiye's 2024 Household Labour Force Survey — 180,908 observations representing 27.7 million private-sector workers across 39 two-digit occupation groups — gives a clear answer for one middle-income economy. Among high-exposure occupations, the largest group is not the one facing displacement. Reorganize accounts for 21.3 percent of private-sector employment, against 12.3 percent in Automation Risk. Reorganization is roughly twice the size of direct substitution risk.
That ratio is the practical finding. Policy built to respond to layoffs will arrive late for most of the workers affected, because the adjustment in Reorganize occupations shows up first as slower hiring, consolidated tasks, and changing hours — while workers are still employed.
Three questions, not one
The framework asks three questions in sequence rather than one.
AI exposure measures how much of an occupation's task content current AI can perform or assist with. This uses the complementarity-adjusted AI occupational exposure index (C-AIOE) of Pizzinelli et al. (2023), in the Turkish mapping developed by Aşık et al. (2026). The high-exposure threshold is the employment-weighted mean across the 39 groups, 4.437, and 41.4 percent of private-sector employment sits above it.
Demand elasticity (η) asks whether AI-driven cost reductions generate enough additional demand to offset the fall in labor needed per unit of output. This is the scale effect, not the wage elasticity of labor demand. Occupation-level estimates of this object do not exist, so they were elicited from standardized occupation profiles — ILO title and definition, five representative tasks adapted to the Turkish private sector, and a description of the associated output — using Claude Sonnet 4.6, three independent runs per occupation, averaged.
Human necessity (H) asks whether delivery still requires a person even when AI can do the component tasks. It is built from O*NET 29.0 through two channels, physical execution and live care and teaching, and mapped to ISCO-08. Long-haul drivers, nurses, and pharmacists score high; data entry clerks and call centre agents score low.
Human necessity is not the same as complementarity. Complementarity asks whether AI makes a worker more productive at a task. Human necessity asks whether the service can be delivered at all without a human in it. An occupation can be highly complementary and still have low human necessity — which is precisely the combination that lands it in Automation Risk.
History supports this distinction. As the US Bureau of Labor Statistics (BLS) recently noted in their AI modeling approach, when digital cameras arrived, they directly stripped the "Human Necessity" from photographic process workers, leading to massive, measurable employment drops. But autonomous driving tech—despite high task "exposure"—hasn't displaced truck drivers because the real-world delivery of the service still requires human oversight. AI is tracking along this exact pattern.
Figure 1. Classification logic. Each occupation group is assigned sequentially: low AI task exposure → Less Change; high exposure + elastic demand → Grow with AI; high exposure + constrained demand + high human necessity → Reorganize; high exposure + constrained demand + low human necessity → Automation Risk. Shares are of total private-sector employment, HLFS 2024, weighted.
The four archetypes
16.2 million workers across 20 occupation groups. Market-oriented skilled agricultural workers (12.9 percent of private employment), drivers (6.0), personal service workers (5.4), labourers (5.0). These fall below the exposure threshold because current AI reaches embodied and situated tasks less directly. Less Change is not AI-proof — it is the archetype most likely to be redefined by robotics.
5.9 million workers across 6 groups. High exposure and constrained demand, but human participation remains integral to delivery. Sales workers alone are 10.6 percent of private employment. Also stationary plant operators, cleaners, food preparation assistants, and legal, social and cultural professionals. The occupation persists; the amount and composition of labor it needs may not.
3.4 million workers across 7 groups. Numerical and material recording clerks (3.5 percent), business and administration associate professionals (2.8), science and engineering professionals (1.9), customer services clerks (1.3), general and keyboard clerks (1.2). Standardized documentation, scheduling, record keeping and reporting, with demand tied to organizational budgets rather than to markets that expand when costs fall.
2.1 million workers across 6 groups, the smallest archetype. Food processing, woodworking, garment and other craft trades alone are just over half of it (4.0 percent of private employment), joined by handicraft and printing workers, business and administration professionals, and ICT professionals and technicians.
What does Reorganize actually look like? The US BLS is currently examining this through targeted occupational case studies in sectors like the legal profession. A lawyer highly exposed to generative AI won't necessarily lose their job (Automation Risk), but the hours previously spent on routine contract review will be reorganized toward client strategy and courtroom presence—shifting the composition of labor without destroying the occupation.
Why sales workers decide the headline
The single most consequential classification in the Turkish application is sales workers, ISCO 52, at 10.6 percent of private-sector employment. Sales is large enough that its assignment moves the aggregate result on its own.
Sales workers have high exposure — AI can support pricing, inventory, customer targeting, transaction processing, and routine service — and high human necessity, since much of the job is presence, interaction, and real-time coordination. The classification therefore turns on demand. Because the occupation spans sectors with very different demand conditions, it is the one group given sector-disaggregated treatment: a sector-weighted elasticity built from the HLFS distribution of sales employment, which is 71.4 percent retail. The resulting estimate, η = −0.85, sits on the constrained-demand side of the −1.0 threshold. Given limited household purchasing power and the necessity-linked character of much retail consumption, cheaper retail operations are not expected to expand the market proportionately.
So sales workers are Reorganize, not Grow with AI. That single call is what makes Reorganize roughly twice Automation Risk, and it is the clearest illustration of the framework's logic: an occupation can be highly exposed, remain thoroughly human, and still face pressure on hours and hiring because demand will not expand to absorb the productivity gain.
A second borderline case runs the other way. Food processing, woodworking, garment and related trades workers (ISCO 75, 4.0 percent of private employment) sit at exactly η = −1.0. The framework's convention is inclusive at the boundary, so they land in Grow with AI — and since they are just over half of that archetype, the entire Grow with AI share depends on a tie-break. Both cases are reported in the paper's threshold sensitivity analysis.
Who is in the exposed archetypes
The aggregate distribution hides substantial variation. These are compositional results: groups differ because they work in different occupations, not because the characteristics themselves cause AI adjustment.
Figure 2. Archetype distribution by gender, educational attainment, and age group, private-sector employment, HLFS 2024. Youth are ages 15–29, prime-age 30–54, older 55 and above.
Women are more exposed than men on both high-exposure margins. Reorganize accounts for 27 percent of female private employment against 22 percent for men, and Automation Risk for 14 percent against 11 percent. Men are correspondingly concentrated in Less Change, 62 percent against 51 percent. The driver is occupational composition: women are heavily represented in retail sales, clerical, administrative, and service work. For women, the more prominent margin is reorganization within continuing roles rather than the occupation disappearing.
The education gradient is the most counterintuitive result. Tertiary-educated workers have the highest Automation Risk share, at 29 percent, and 22 percent in Reorganize — against 70 percent Less Change among workers with below-secondary education. Higher education does not insulate workers here, because in Türkiye the tertiary-educated are disproportionately in formal cognitive, clerical, administrative, and analytical occupations, which is exactly where current AI capability lands. (This isn't just a quirk of the Turkish labor market. When the US BLS selected occupations for its first wave of AI-impact case studies, they bypassed the factory floor entirely, focusing instead on computer, legal, business, and engineering professionals—the exact tertiary-educated cohorts our framework flags.) Tertiary workers also have the largest Grow with AI share, at 11 percent. Education does not map onto a single adjustment mechanism; it determines which exposed occupation a worker is in.
Age differences are milder. Youth have a somewhat larger Automation Risk share than prime-age workers, at 15 against 12 percent, consistent with entry-level clerical and administrative roles. Reorganize is about 24 percent for both. Older workers are heavily in Less Change, 76 percent, which reflects where they work rather than any AI-complementarity advantage.
Where the exposure sits
Figure 3. Archetype distribution across NACE sectors, private-sector employment, HLFS 2024. Sectors accounting for at least 2 percent of private employment; sorted by Reorganize share. These are occupation-composition results — sectors differ because they employ different mixes of occupations, not because sector-level AI adoption is measured.
Trade is the clearest Reorganize sector, at 57 percent of its employment, driven by the concentration of sales workers in retail. Trade is 17 percent of private-sector employment, so this alone is consequential for the aggregate. Administrative and support services combines 19 percent Automation Risk with 33 percent Reorganize, mixing clerical functions with cleaning and coordination work. Accommodation and food services is 32 percent Reorganize, which shows that reorganization extends well beyond office work into human-intensive service delivery.
Manufacturing, at 23 percent of private employment, is the most internally mixed: 25 percent Reorganize alongside 21 percent Grow with AI, reflecting a workforce of machine operators, assemblers, craft workers, and technical roles that the framework routes down different branches. Professional services is the most exposed sector overall — 38 percent Automation Risk, 25 percent Reorganize, 17 percent Grow with AI, and only 19 percent Less Change — and illustrates the central point that high exposure does not imply one adjustment mechanism. Agriculture, 17 percent of private employment, is classified entirely as Less Change.
Figure 4. Archetype distribution across NUTS-2 regions, labeled by lead province, sorted by Reorganize share. Private-sector employment, HLFS 2024.
The regional pattern is metropolitan and industrial. İstanbul — 22 percent of private-sector employment on its own, 6.0 million workers — combines the largest Reorganize share, 28 percent, with the largest Automation Risk share, 18 percent. Tekirdağ matches it on Reorganize; Ankara and Bursa follow at 25 percent, then Gaziantep, Kocaeli and İzmir at 24. At the other end, Ağrı is 85 percent Less Change and Van 78 percent.
Lower exposure in the eastern and agricultural regions should not be read as preparedness. Those labor markets combine less exposed occupations with weaker access to training, employment services, digital infrastructure, and social insurance. If AI capability extends into physical and embodied tasks, the regions with the most time to prepare are the ones least equipped to use it.
Formality: the exposed workers are the registered ones
Figure 5. Archetype distribution by formality status and establishment size, private-sector employment, HLFS 2024. Formal is defined as social-security registration in the main job. Establishment size categories are one-person, micro (2–9), small (10–49), and medium-to-large (50+).
Formal workers are far more concentrated in the exposed archetypes than informal workers. Reorganize is 26 percent of formal employment against 15 percent of informal; Automation Risk is 15 percent against 3 percent. Informal workers sit overwhelmingly in Less Change, at 76 percent, because informal employment in Türkiye is concentrated in agriculture, construction, household services, and transport — physical and situated work that current AI reaches less directly.
This inverts a common assumption. Informality is usually treated as the vulnerable margin of a middle-income labor market; under current AI capabilities it is the less exposed one. But lower exposure is not resilience. Informal workers have weaker access to employer-provided training, public employment services, and social insurance — so if their occupational exposure rises, they face the adjustment with fewer institutional supports, not more.
Establishment size shows the same logic from another angle. Automation Risk rises monotonically with size, from 7 percent in one-person establishments to 19 percent in establishments of 50 or more, because larger firms employ more administrative, analytical, and coordination staff. One-person establishments are 65 percent Less Change.
The View from the US: While the Richmond framework uses predictive modeling to forecast structural changes before they happen, the US Bureau of Labor Statistics takes a strictly empirical approach, waiting for data trends to materialize before adjusting their employment projections. Yet, both approaches arrive at the same conclusion: task evolution (like the adoption of smartphones or LLMs) rarely equals immediate job destruction. The transition is about reorganization, not just replacement.
Türkiye against the United States
Richmond's (2026) U.S. application provides a reference point, though the two differ in exposure measures, occupational classification, employment coverage, elasticity inputs, and aggregation — so the comparison is interpretive rather than a harmonized decomposition.
| Archetype | Türkiye | United States |
|---|---|---|
| Less Change | 58.7% | 46% |
| Reorganize | 21.3% | 24% |
| Automation Risk | 12.3% | 18% |
| Grow with AI | 7.7% | 12% |
The comparison does not support the presumption that a middle-income economy carries more automation risk. Türkiye's Automation Risk share is lower than the United States', 12.3 against 18 percent, because so much of its private employment remains in agriculture, construction, transport, manual trades, and personal services. Reorganize is the most similar category across the two, 21.3 against 24 percent, and is the largest high-exposure archetype in both.
The distinctive Turkish features are elsewhere: a substantially larger Less Change share, 58.7 against 46 percent, and a smaller Grow with AI share, 7.7 against 12 percent. The second is the one that should worry policymakers. It says that under the implemented classification, fewer high-exposure occupations in Türkiye meet the demand-elasticity criterion — the channel through which productivity gains turn into employment growth is narrower.
What follows for policy
Because Reorganize is roughly twice the employment share of Automation Risk, the central institutional task is to detect restructuring before it surfaces as unemployment or loss of formal employment.
- Automation Risk occupations need transition pathways, and early monitoring. Adjustment here is likely to appear first as weaker recruitment, hiring freezes, and reduced demand for administrative support rather than abrupt layoffs. Monitoring should focus on entry-level vacancies and hiring into clerical and administrative occupations. These workers hold transferable organizational, communication, and coordination skills; training should be tied to identified adjacent roles rather than to generic digital skills.
- Reorganize occupations need in-work support, not post-displacement support. Adjustment begins while workers are still employed, through task allocation, hours, advancement, and hiring. That means in-work adaptation training, sector-specific monitoring of hours and staffing in retail, hospitality, food services, administrative support and building services, and continuity of social protection as employment arrangements shift. Policies triggered only by job loss respond too late.
- Grow with AI is a demand-side problem, not only a skills problem. The archetype is small because few high-exposure occupations clear the elasticity bar. Widening it means reducing the constraints — fixed costs, information, managerial capability, process reorganization — that limit firms' use of AI-complementary professional, technical, and specialized production services.
- Less Change warrants preparedness, not complacency. These occupations are concentrated among informal workers, one-person establishments, and lower-exposure regions — exactly the groups with the weakest access to training, employment services, and social insurance. Advances in robotics and embodied AI could move this boundary while institutional capacity is still thin.
Across all four, the same three institutional functions recur: early-warning systems that track hiring, hours, separations and task requirements rather than only unemployment; anticipatory employment services that reach workers while they are still employed; and continuity of social protection as workers move between employers and employment arrangements.
What this does not claim
The analysis is diagnostic, not causal or predictive. It does not measure actual AI adoption, model the timing of technology use, or forecast job losses. An occupation classified as Automation Risk is not predicted to lose a given number of jobs.
Three limitations matter most. The classification runs at the ISCO-08 two-digit level, which masks differences in task content across workers, firms, and regions within an occupation group — a caveat that applies with particular force to internally heterogeneous groups such as science and engineering professionals. The demand-elasticity inputs are elicited from standardized profiles rather than estimated from observed market responses, so they depend on profile design, prompt, model version, and the Türkiye-specific assumptions supplied. And the human-necessity scores are derived from U.S. O*NET data mapped to ISCO-08, with activation thresholds informed by prior occupational information rather than estimated from Turkish data.
The framework is also partial equilibrium: no wages, no endogenous task creation, no firm entry and exit, no sectoral reallocation, no economy-wide feedback. And because the classification is threshold-based, occupations near the cutoffs — sales workers and craft trades above all — can move between archetypes under alternative specifications. Less Change remains the largest archetype under every threshold tested, and Reorganize remains larger than Automation Risk when the exposure and elasticity thresholds are varied, but the ordering does reverse under a stricter human-necessity threshold.
Subject to those limits, the point stands. Exposure tells you where AI reaches. Adding demand responsiveness and human necessity tells you what reaching there is likely to do — and in Türkiye, the answer for most exposed workers is reorganization rather than displacement.
Comments
Post a Comment
Constructive feedback is always welcome. Thank you