Working paper · August 2026 draft
Generative AI and Labor Displacement in the Global South: Evidence from Philippine Business Process Outsourcing
Nico Ravanilla
In brief
Generative AI can do the work the Philippine outsourcing industry sells. Using the Anthropic Economic Index and Philippine labor force data, this paper finds employment growth has already slowed in the most AI-exposed provinces, industries, and municipalities, amounting to roughly a quarter million foregone jobs by mid-2025, with the adjustment falling on hiring rather than wages, and with overseas work shrinking rather than absorbing the workers the sector no longer hires.
Abstract
Exported services have become for labor-abundant countries what manufacturing once was: a development path that absorbs large numbers of workers into internationally traded production. Generative AI performs many of the tasks these economies export, raising the possibility of labor displacement. I examine whether this displacement has begun in the Philippines, the economy most dependent on this path, by combining pre-ChatGPT occupational composition with occupation-level observed AI usage from the Anthropic Economic Index and tracing labor market outcomes as US client firms adopt AI. Employment growth slows in the most exposed provinces, industries, and municipalities, amounting to roughly a quarter million foregone jobs by mid-2025. The adjustment occurs through reduced hiring rather than layoffs: wages and hours do not change. The displaced enter unemployment, and first-time overseas departures fall rather than absorbing the workers the sector no longer hires.
Research annotationQuestion, variables, design, findings, and mechanism at a glance▾
- Research question
- Has generative AI begun displacing labor in the Global South, and on which margin does the adjustment appear first?
- Independent variable (X)
- Sub-national AI exposure interacted with the US adoption dose. Exposure combines pre-ChatGPT occupational composition with occupation-level observed AI usage from the Anthropic Economic Index, so it is fixed before the shock. The dose is the share of US firms using AI in production from the Business Trends and Outlook Survey, entering at a one-quarter lag because client adoption reaches Philippine payrolls through hiring freezes, attrition, and contract renewals. Demand is American, so US adoption rather than Philippine usage measures the shock.
- Dependent variable (Y)
- BPO employment share and log employment, daily pay and weekly hours, the occupation mix within BPO, employment in nine destination sectors, regional unemployment and labor force participation, and first-time versus repeat overseas deployments.
- Identification strategy
- Three designs on Philippine Labor Force Survey panels covering 87 provinces over 17 quarterly rounds through 2025Q2, all including exposure-specific linear trends so the coefficient is identified by the curvature of the adoption path rather than by any confound trending smoothly with exposure. The province design regresses outcomes on provincial exposure interacted with the lagged dose. The within-province design replaces it with industry-level exposure and adds province-by-round and province-by-industry fixed effects, so identification comes from differences across industries inside the same province and quarter, absorbing every province-level shock. The municipal design interacts a municipality's pre-2022 IT-zone Presidential Proclamations with the dose, with municipality and province-by-round fixed effects. Regressions are weighted by pre-period employment and clustered at the treated unit. Pre-release placebo and curvature tests pass for the within-province and municipal designs but not for the province design, whose pre-period spans the capital's pandemic reopening, so the province estimates carry descriptive weight while the other two carry the identification. The overseas margin uses a province-by-quarter panel of Department of Migrant Workers deployment records.
- Main findings
- Employment falls across all three designs. A province one standard deviation more exposed loses 0.149 percentage points of BPO employment share (SE 0.027) and 3.7 percent of BPO employment (SE 1.2) per adoption point; within provinces an industry one standard deviation more exposed loses 0.6 percent weighted and 1.2 percent unweighted; municipalities with more IT-zone infrastructure lose 0.060 log points (SE 0.026). Cumulated over the observed adoption path this is a 16 log point shortfall by mid-2025, or roughly 240,000 foregone jobs, about one eighth of the sector. Pay and hours do not respond, with the pay coefficient of 0.001 (SE 0.004) ruling out wage effects larger than one percent per adoption point. The response peaks one to two quarters after adoption and attenuates by the fourth. No destination sector absorbs the displaced: every coefficient lies within 0.02 log points of zero, and the total employment effect of -0.0013 (SE 0.0020) is close to the -0.0015 benchmark for complete displacement without reallocation. Regional unemployment rises 0.29 percentage points per adoption point per unit of exposure (SE 0.04, wild cluster bootstrap p = 0.013) while participation does not move. First-time overseas deployments fall 2.0 percent per adoption point per exposure standard deviation (p = 0.004), while rehires do not respond.
- Mechanism
- Displacement operates at the hiring margin. Firms stop adding workers as clients automate the tasks they previously bought, rather than cutting pay or separating incumbents, and within BPO the occupation mix rotates away from high-exposure occupations by about 7 percent per exposure standard deviation per adoption point, consistent with incumbents being reassigned while entry closes. For a workforce this young, foregone entry is where displacement first appears. The fall in first-time overseas departures is broad-based across age, sex, and destination, which points to a shock to the household resources that finance migration rather than to any particular corridor.
- Why it matters
- Provides some of the first sub-national causal evidence on generative AI's labor market effects outside rich countries, in the industry that is among the developing world's largest employers and most directly exposed. Its sharpest implication is that the two conventional escape valves close together: displaced workers are not absorbed by other sectors and the overseas labor market, historically the Philippines' principal adjustment margin, contracts rather than expands, so exported services may not carry labor-abundant economies the way manufacturing once did.
- Speaks to
- AI and labor marketstask-based and routine-biased technological changetrade in services and offshoringmigration and remittancesdevelopment economics
Cite
@misc{ravanilla_ai_bpo,
title = {Generative AI and Labor Displacement in the Global South: Evidence from Philippine Business Process Outsourcing},
author = {Nico Ravanilla},
howpublished = {Working paper},
}