TL;DR: A bunch of emerging economies built their development models around service-sector offshoring. AI is collapsing the cost arbitrage that made that work. The Philippines has over a million BPO jobs with immediate exposure, and the fiscal capacity to manage the transition isn't there. New jobs will come eventually, but the gap between displacement and creation is the crisis. I lay out four policy mechanisms and explain why none of them work without coordination.

Adapted from research conducted during graduate studies at Columbia University, 2023-2025.

Here's the core tension: over the past two decades, a bunch of emerging economies bet their development strategies on service-sector offshoring. The Philippines built 10-15% of GDP around BPO. Indian IT services became a $200+ billion industry. South Africa’s service sector hit 70% of GDP. Rodrik called it “premature tertiarization” and worried about it, but the numbers looked good — growth accelerated, formal employment expanded, development indicators improved.

The bet worked because of a simple arbitrage: cognitive labor in Manila or Hyderabad was cheaper than in New York or London, and telecom infrastructure made it deliverable at distance. Thirty years of global value chain fragmentation, the phenomenon Baldwin documented so thoroughly, made this possible.

Now generative AI is collapsing that arbitrage. The same characteristics that made service offshoring viable—routine cognitive tasks, digital deliverability, labor cost gaps—are precisely what make these sectors maximally exposed to automation. This isn’t a parallel to manufacturing automation in the Rust Belt. It's potentially the unwinding of an entire development model — and it's hitting economies with far less fiscal capacity to manage the transition.

I want to be precise about where my confidence is high and where I'm speculating.

Three Properties That Make This Different

Most of the automation literature we lean on—Acemoglu and Restrepo on robots, Autor’s task framework—draws from manufacturing contexts. AI's economic properties differ in ways that matter for the Global South.

The arbitrage collapse. Traditional automation had high fixed costs and positive marginal costs. Capital constraints limited deployment speed. AI automation has high upfront costs in model development but near-zero marginal costs for additional use. Once a frontier model exists, serving one more query costs fractions of a cent.

The traditional offshoring arbitrage worked because:

(Cost of offshore labor) < (Cost of domestic labor) - (Coordination costs)

AI changes this to:

(Cost of AI) < (Cost of any human labor, regardless of location)

And coordination costs now favor proximity. AI systems integrate more easily with local operations than distant offshore teams. The economic logic that sustained three decades of offshoring is reversing. Klarna reshored customer service from Manila to Stockholm-based AI systems. India’s IT sector job posting growth fell 30% year-over-year per NASSCOM’s 2024 data. This is structural, not marginal.

Task substitution without complementarity. The ALM framework told us automation handles routine tasks but complements workers on non-routine ones, allowing reallocation within occupations. AI breaks that pattern. It automates routine tasks, yes, but it’s increasingly capable of non-routine cognitive work too. The complementarity still exists, but for a much smaller slice of the workforce.

Consider customer service. In 2010, IVR systems filtered volume and human agents handled everything else—automation complemented the workforce. In 2025, LLMs handle simple and moderately complex queries, and only the highest-complexity cases need humans. AI substitutes for roughly 80% of the workforce. Only the top 20% see complementarity benefits.

This creates what I think of as skill premium inversion: previous automation waves raised returns to college education broadly. AI may actually reduce returns to bachelor’s-level cognitive work while dramatically increasing returns only to advanced expertise combined with AI fluency. We’re seeing inequality within skill groups in ways that don’t map onto any historical automation episode I’m aware of. The workers in Global South service sectors sit disproportionately in that bachelor’s-level band.

Diffusion speed as the binding constraint. Industrial robots went from invention in the 1960s to widespread use in the 1980s-90s. PCs took a similar arc. AI is compressing this to 2-5 years: GPT-3 in 2020 to widespread enterprise deployment by 2024-25. Traditional labor market adjustments assumed 10-20 year windows. Educational systems could adapt. Safety nets could be built. AI may give these economies 3-5 years.

You could argue diffusion will be slower than I’m projecting. Enterprise adoption is messy: integration challenges, organizational inertia, regulatory friction. Fair. But even if I’m off by a factor of two and it takes 6-10 years, that’s still dramatically faster than any previous general-purpose technology wave, and still faster than the institutional response capacity of most emerging economies. Better to build policy for the faster scenario and be pleasantly surprised.

The Philippines Problem

Generic claims about AI displacement (“300 million jobs at risk”) operate at the wrong level of abstraction. Which tasks, when, and in which economies?

The Philippines is the sharpest case of what I call Archetype A: high-concentration service economies (Jamaica fits here too, as do specific Indian metros like Bangalore and Hyderabad). Services exceed 60% of GDP. BPO alone is 10-15%. Metro Manila accounts for 60% of the country’s BPO employment. It’s sectoral monoculture, and monoculture creates systemic risk.

The numbers are sobering. Tier 1 BPO support (password resets, account inquiries, basic troubleshooting) employs roughly 700,000 people in the Philippines. Automation feasibility for these tasks is already at 95%, demonstrated at scale. Klarna is handling two-thirds of customer support volume with AI, equivalent to 700 agents. Content moderation (another 150,000 jobs) is already being automated by Meta, TikTok, and YouTube. Add basic data entry and you’re looking at over a million jobs with immediate exposure, at wage levels of $3-15/hour against AI equivalent costs of $0.50-2/hour.

Model a 40% BPO employment decline over three years. Direct job losses: 700,000. Apply Acemoglu et al.’s multiplier estimates of 1.5-2x, and total unemployment impact reaches 1-1.4 million. GDP impact: 3-5%. The Philippines has debt-to-GDP at 60% and social spending at 2-3% of GDP. It can't afford Danish-style safety nets. The fiscal capacity simply isn’t there.

The spectrum runs from there to diversified emerging economies like India or Brazil at the national level—large domestic markets provide a buffer, manufacturing coexists with services, and greater fiscal space exists—down to manufacturing-focused late industrializers like Vietnam and Bangladesh, which face different automation dynamics on a different timeline. But the Philippines archetype is where urgency is highest. Applying Hausmann-Rodrik-Velasco’s growth diagnostics framework, the binding constraint in these economies isn’t identifying the problem. It’s that the standard toolkit (diversification, institution-building, education reform) operates on timescales of 5-20 years, and they may have 3-5.

“AI Will Create New Jobs We Can’t Predict”

This is the strongest counterargument, and it deserves a real answer. Every previous technology wave (mechanization, electrification, computing) destroyed jobs and created new ones we couldn’t have anticipated. Autor’s own work shows the U.S. economy generating entirely new job categories that absorbed displaced workers. Why should AI be different?

Here’s the steelman: AI is a general-purpose technology. GPTs historically create new industries, new forms of demand, new occupations. The pessimistic case requires believing “this time is different,” which is exactly what people wrongly believed about ATMs, spreadsheets, and the internet. Technological unemployment has been predicted and failed to materialize for 200 years. The base rate for “technology permanently reduces employment” is zero.

The historical record supports this. But the speed and marginal-cost properties do make this time structurally different, at least for the transition period that matters for policy.

Previous GPTs had long diffusion periods (decades) during which labor markets adjusted. New industries had time to form. Education systems adapted. Workers retrained across cohorts, not within careers. AI is compressing this to years. And the near-zero marginal cost property means that once a task is automated, there’s no economic floor under which human labor becomes competitive again. With manufacturing automation, different geographies could compete on labor costs. When AI costs fractions of a cent per query, no amount of wage flexibility makes human workers competitive for the automated tasks.

The new jobs will come. I believe that. But they’ll come on a 10-20 year horizon, and the displacement is happening on a 3-5 year horizon. The gap between destruction and creation is the crisis. For the Philippines, that gap could mean a lost decade of development progress. The long-run equilibrium is probably fine. The transition path is where economies with weak institutional capacity break.

The Policy Trilemma

Policy responses face a trilemma between speed, equity, and efficiency. AI displacement operates on 2-5 year timescales. Skills retraining takes 1-3 years per cohort. Economic diversification takes 5-10 years minimum. You can't build adaptive capacity fast enough through traditional means.

Meanwhile, AI productivity gains concentrate at the top, so market-based transitions dramatically increase inequality. But slowing adoption reduces growth, and labor protections create distortions that international competition punishes.

You can optimize for two dimensions, not three.

Speed + Efficiency (sacrifice equity): Rapid adoption, market adjustment, minimal intervention. BPO sectors shrink 50% in three years, new sectors emerge but employ fewer, inequality spikes 8-10 Gini points. Growth continues but gains concentrate. Social instability risk is real.

Speed + Equity (sacrifice efficiency): Aggressive redistribution, heavy taxation of AI gains. Requires 3-5% of GDP in transfers. The Philippines lacks the fiscal capacity. Possible only with massive international financing and likely creates market distortions.

Equity + Efficiency (sacrifice speed): Regulate AI deployment, coordinate transition. But if the Philippines slows adoption, work shifts to India. If India regulates, work shifts to Vietnam. Unilateral regulation is unstable. You’d need international coordination, and I’m not optimistic about that.

The trilemma is real. The question is which trade-offs to accept.

Four Policy Mechanisms Worth Taking Seriously

Moving past “invest in education” and “strengthen safety nets”—four mechanisms addressing specific market failures, designed with emerging economy constraints in mind.

Conditional AI deployment licensing. Firms don’t internalize the social costs of rapid displacement. Require firms deploying AI systems that displace more than 50 workers to submit transition plans with minimum timelines, retraining commitments, and severance requirements. This is a Pigouvian response to a real externality—it slows displacement enough to enable adjustment while preserving incentives to adopt AI. Firms accept it to avoid harsher regulation. The precedent exists: the U.S. WARN Act, German Works Councils. For the Global South, simplify compliance requirements, phase in starting with large firms, and harmonize regionally to prevent arbitrage.

AI productivity levy for worker transition. AI productivity gains concentrate among capital owners; displaced workers capture none. Index a small levy (2-5% of documented labor cost savings) on AI deployment and channel it to a Worker Transition Fund — financing things like wage insurance, training subsidies, relocation assistance, and entrepreneurship grants. It’s self-financing: AI savings fund the transition. In the Philippines, a 5% levy on estimated BPO wage savings would generate roughly $750 million over four years, enough to support 400,000 workers. The challenge is measurement and enforcement; the mitigation is to use deemed savings based on industry benchmarks rather than firm self-reporting.

Skills credential bonding. Training programs have dismal placement rates (often below 40%) yet receive public funding without accountability. Create outcome-based credential bonds: training providers issue bonds that pay out contingent on graduate employment within six months. Bonds are tradeable, creating price signals about program quality. Providers with poor placement rates face market discipline. Government subsidizes the bond premium to reduce student risk. This aligns incentives in a way that traditional education subsidies don’t.

Regional economic diversification bonds. Individual firms won't invest in regional diversification — coordination failures. Issue place-based development bonds—say, $500 million for a solar manufacturing cluster in Metro Manila—with private co-investment commitments contingent on infrastructure completion. Bonds pay premium only if employment targets are achieved. This solves the coordination failure by getting multiple firms to move together, with risk-sharing between public infrastructure investment and private operations. Blended finance from development banks makes the capital stack feasible.

The Coordination Problem

Every one of these mechanisms is weakened by a collective action problem. AI productivity benefits are global (the technology diffuses). Displacement costs are local (concentrated geographically). Countries compete for AI investment, creating a race-to-bottom on worker protections. If the Philippines implements deployment licensing that costs firms 10% of savings, firms credibly threaten relocation to Vietnam.

What currently exists—OECD principles, UNESCO recommendations, the EU AI Act—is either non-binding, regional in scope, or unenforceable. What we actually need: binding commitments on transition support, harmonized minimum standards, and financing mechanisms for Global South adaptation.

The realistic path is regional harmonization—ASEAN countries coordinating on AI labor standards, reducing competitive pressure within the bloc. Multilateral development banks tying lending to transition policies would help too. Neither is fast. Global coordination without a crisis catalyst is unlikely, and the crisis may get there first.

What I’m Confident About, and What I’m Not

High confidence: current AI systems can automate specific BPO and IT tasks at scale—that’s demonstrated, not projected. Adoption is measurably occurring. Some displacement is already visible in job posting declines and headcount reductions. Historical automation also consistently increased inequality absent policy intervention.

Medium confidence: displacement will accelerate significantly in the 2-4 year timeframe, that reshoring will increase as AI costs fall, and that country-level GDP impacts will range 1-5%. The economic logic is clear even where direct evidence is still emerging.

Low confidence: AI capability trajectories beyond three years, the magnitude and type of new job creation, political system responses, and the likelihood of meaningful international coordination. Extrapolating beyond a few years is speculation, and worth being upfront about that.

The cost of over-preparing is investment in safety nets, skills development, and economic diversification — things with positive returns regardless. The cost of under-preparing is a lost generation of development progress.

The Clock

What makes AI different as an economic force is the combination of features: faster diffusion than any previous GPT, near-zero marginal costs, cognitive rather than manual task automation, and services rather than manufacturing as the primary locus of disruption. These properties don’t fit cleanly into existing analytical frameworks—not Autor’s task model, not Acemoglu-Restrepo’s automation theory, not standard trade models of offshoring. We need new theory informed by, but not constrained by, historical patterns.

Can the Global South navigate this without losing a decade of progress? Yes, but it requires acknowledging uncertainty, evidence-based policy, political courage to act before the crisis forces action, and international cooperation where interests align. History says reactive is more common. Economics says proactive is cheaper.

I estimate 2-3 years before displacement outpaces adaptation capacity in high-exposure economies. You could argue I’m being aggressive—institutional friction, procurement cycles, regulatory drag all buy time. Maybe. But the cost of being wrong in the cautious direction is higher than the cost of being wrong in the urgent direction. After that window closes, we’re in crisis management, not strategic transition.

The stakes are hundreds of millions of workers and their families. The window is narrow and closing.

Nikhil Ghosh conducted this research during graduate studies at Columbia University, 2023-2025. Contact: nrg2156@columbia.edu

Things I’ve Been Reading