Why does AI belong in development policy?
Before asking whether a country needs an AI bank, there is a prior question. Why should artificial intelligence be treated as a development transition at all?
Forecasts of mass unemployment are too uncertain to carry the argument. No credible source can yet say how many jobs will disappear, how quickly firms will reorganize or how much of the technology's capability will translate into productivity. The estimates are too sensitive to assumptions about adoption, task redesign, regulation and complementary investment.
The reason for government attention is the combination of scale, uncertainty and uneven exposure. AI can alter production across sectors, reward firms and countries that possess compute, data and skills, and weaken industries elsewhere before those economies have built new sources of income. Its gains are unlikely to arrive in the same places, or at the same time, as its losses.
This is where the comparison with climate finance begins. Climate change forced governments to look beyond a collection of environmental projects and confront an economy-wide transition. The public task included infrastructure, standards, disclosure, skills, industrial change, risk allocation and support for communities carrying the cost. Finance turned distant targets into decisions about real assets and institutions.
AI is a different problem, but it has the same institutional reach. Electricity and data centres matter. So do cloud access, local-language data, company systems, management practice, research, public procurement, competition, taxation and worker adjustment. A national AI plan that does not specify how these capabilities will be financed is closer to an aspiration than a transition strategy.
The case does not begin with a preferred bank. It begins with a development risk: a country can experience AI disruption at global speed and capture the productivity dividend only at the speed of its own institutions.
Disruption can arrive before the dividend
Published estimates of AI's economic effect span an unusually wide range. UNCTAD projects the AI market itself to grow from $189 billion in 2023 to $4.8 trillion in 2033. That is a market forecast, not a prediction of additional output. WTO simulations find that AI could leave global trade 34 to 37 percent higher and GDP 12 to 13 percent higher by 2040, depending on technology costs and adoption. At the cautious end, Daron Acemoglu estimates that currently measurable task exposure implies no more than a 0.66 percent increase in total factor productivity over ten years. OECD estimates for the G7 sit between those poles.
These figures cannot be averaged into a reliable number. They describe different concepts and scenarios. Their spread is itself useful. Governments are facing a technology with potentially large macroeconomic consequences and no settled path of diffusion.
The distribution is easier to see. In 2022, 100 firms, concentrated mainly in the United States and China, accounted for 40 percent of global corporate AI research and development. The same two countries held 60 percent of AI patents. UNCTAD found 118 countries, mostly developing economies, absent from the principal international AI-governance initiatives it reviewed.
Labour exposure follows another uneven pattern. The IMF estimates that almost 40 percent of employment worldwide is exposed to AI: about 60 percent in advanced economies, 40 percent in emerging markets and 26 percent in low-income countries. Exposure includes work AI may complement and work it may substitute. The ILO's 2025 index finds one worker in four in an occupation with some exposure, with 3.3 percent of global employment in the highest-exposure category. Clerical work remains the most exposed, but professional and technical roles are moving into range.
Lower measured exposure is not protection. A country can have fewer automatable jobs and still fall further behind if firms elsewhere raise productivity more quickly. It can also lose an export niche without automating much of its own economy. Customer support, back-office processing, software services and routine design work can be displaced through purchasing decisions made by clients abroad.
The World Bank and ILO call this “disruption without dividend”. In many developing economies, workers in automation-vulnerable activities are connected enough to face displacement. Workers whose jobs could become more productive may be in firms without reliable connectivity, data, digital systems or managerial capacity. The first group can encounter AI through foreign competition before the second encounters it as a useful tool.
Select a position in the AI economy. These are analytical types, not permanent country labels.
Possible dividend
AI can augment workers, defend contracts and support movement into higher-value services.
Transition risk
Foreign clients can automate demand before domestic firms and workers have built the next export industry.
Public-finance priority
Compress the transition: fund adoption, supplier upgrading, portable skills and new export capabilities while protecting fiscal room.
The Philippines is a stress test
The Philippines shows how this asymmetry could work. Its business-process outsourcing industry was built by connecting a large English-speaking workforce to foreign demand. According to an IMF study of the Philippine labour market, the sector generated $35.5 billion in revenue in 2023. It employed only about 3 percent of the workforce but produced income equivalent to roughly 7.4 percent of GDP, close to the value of remittances. North America supplied about 70 percent of its market.
The same study estimates that 36 percent of Philippine employment is highly exposed to AI. It places 22 percent in high-exposure occupations where AI is more likely to complement workers and 14 percent in high-exposure occupations with lower complementarity. Customer-service representatives, telemarketers, accounting staff, auditors, secretaries and administrative clerks appear among the vulnerable groups.
| Source | Measure | Philippine estimate |
|---|---|---|
| IMF, 2025 | High AI exposure | 36% of employment: 22% higher-complementarity; 14% lower-complementarity |
| ILO, 2026 | Any generative-AI exposure | 12.7M workers; 3.6% of employment in the highest-exposure category |
| World Bank, 2026 | High exposure, low complementarity | Roughly one-fifth of workers under its methodology |
The ILO and a separate World Bank assessment produce different numbers because they ask different questions. They do not know how quickly clients will adopt, how firms will reorganize or whether cheaper AI-assisted delivery will expand demand. Philippine providers can use AI to improve quality and retain contracts. The country also has higher-value niches in health information, analytics and compliance.
Foreign client
Chooses whether to buy Philippine labour, AI-assisted service or a largely automated platform.
Philippine economy
Receives export earnings, wages, spending and payroll-linked revenue when the work remains.
AI provider
Can capture subscription revenue, intellectual-property rents and company value in another jurisdiction.
The downside channel is clear. A foreign client that replaces routine contact-centre work with an AI service can reduce Philippine export receipts, wages, consumption and payroll-linked revenue. The corresponding profit may accrue to a model provider or cloud platform headquartered elsewhere. The Philippines can tax local consumption of digital services and may secure some withholding or corporate revenue, but that is not the same fiscal base as a large domestic payroll and locally distributed business income.
There is not yet strong causal evidence of an AI-driven collapse in Philippine outsourcing. Treating it as inevitable would be irresponsible. Treating the exposure as somebody else's commercial decision would be equally complacent. It is a plausible, nationally significant risk with a lead time for investment that may be shorter than the lead time for retraining millions of workers or building new export industries.
A post-work tax base does not travel automatically
American discussion of advanced AI often jumps from mass automation to abundance and redistribution. In his 2021 essay Moore's Law for Everything, Sam Altman proposed an American equity fund capitalized by taxes on large companies and private land, with annual distributions to every adult. The premise was that AI-generated wealth could be captured from appreciating American corporate assets and shared with American citizens.
That argument contains a geographic assumption. The companies, their equity and a large part of their taxable value sit within reach of the state paying the dividend.
For a country that mainly consumes AI services, the fiscal arithmetic can run in the opposite direction. An IMF scenario exercise published in 2026 warns that falling employment and wage income could erode personal-income and payroll tax bases. Higher corporate tax receipts may only partly offset the loss because capital is mobile, market power is concentrated and countries compete over tax. The IMF says the problem is most acute for economies that consume AI services rather than produce them. It treats UBI-type programs as an option under an extreme rapid-displacement scenario, not as an automatic dividend from technical progress.
The international tax system reinforces the asymmetry. IMF work on cross-border services notes that a remote foreign supplier can earn substantial revenue from a country while leaving it limited rights to tax the associated income under traditional rules. VAT, withholding taxes, digital-services taxes and expanded concepts of taxable presence can recover part of the base. They do not ensure that a country losing a labour-intensive export industry captures the rents earned by the technology replacing it.
The policy objective is to shorten the interval between exposure and capability: help firms adopt and adapt the technology, build new sectors, finance portable skills and ensure that tax and social-protection systems can survive the adjustment.
Ordinary labour policy is not enough. A training budget responds after a skills gap is identified. A transition strategy may need to finance the systems that make new work possible: reliable power, cloud access, shared compute, company digitization, local data, applied research, testing facilities, supplier upgrading and new export platforms. Some are public goods. Some are investable assets. Some sit inside firms that cannot offer conventional collateral.
What the climate-finance comparison offers
Climate and AI are not equivalent crises. Climate policy responds to a physical externality and an agreed need to reduce emissions and strengthen resilience. AI policy concerns a privately developed general-purpose technology whose net effects can be productive, disruptive or both. The climate label cannot justify indiscriminate subsidy, technological nationalism or a public institution with no viable assets to finance.
The useful comparison is about the structure of transition.
| Dimension | Climate transition | AI transition | Institutional consequence |
|---|---|---|---|
| Reach | Energy, transport, buildings, land and industry | Production, services, research and public administration | Coordination extends beyond one ministry or lender |
| Complementary capacity | Grid, standards, project preparation and skills | Power, connectivity, compute, data, systems and skills | Capital alone cannot produce the transition |
| Distribution | Stranded assets, regional and worker adjustment | Exposed tasks, export industries and mobile technology rents | Finance must sit beside labour, competition and fiscal policy |
| Financeability | Many long-lived physical assets and measurable outputs | A mix of infrastructure, fast-depreciating hardware, intangibles and public goods | An AI institution needs a stricter asset and instrument test |
Both transformations cut across sectors and depend on infrastructure and capabilities that no single firm will build in full. Both can produce cumulative advantage: early infrastructure, standards, skills and supplier networks attract more investment. Both distribute costs and benefits unevenly across regions, industries and workers. Both require long-term coordination while technology and economics are changing. In both cases, a financing decision can commit the economy to a path long before the final outcome is known.
Climate finance also offers a warning. The response did not consist of creating green banks everywhere. In the years around the Paris Agreement, the international community built a much wider architecture.
The Green Climate Fund mobilizes $9.3 billion
A large multilateral pool gives developing countries access to concessional finance for mitigation and adaptation.
The Green Climate Fund mobilized $9.3 billion in 2014 for mitigation and adaptation in developing countries. The Financial Stability Board created the Task Force on Climate-related Financial Disclosures; its 2017 recommendations made climate risk visible to investors and lenders. Central banks and supervisors formed the Network for Greening the Financial System. The European Union developed taxonomies, disclosure rules, benchmarks and investor duties. Labelled green-bond issuance grew from $3.4 billion in 2012 to $167 billion in 2018.
These initiatives addressed different failures: scarce concessional capital, weak project pipelines, inconsistent definitions, opaque risk and limited investor demand. Within that system, the green investment bank came to the fore as one specialist tool. The OECD's 2016 study examined more than a dozen public banks and bank-like entities designed to mobilize private capital for domestic low-carbon and climate-resilient investment.
The model created a practical choice. If a government already owned a development bank with capital, clients and operating systems, should it add a climate mandate or create a specialist institution?
The choice Smallridge et al. tested
Diana Smallridge, Marta Becker, Jenni Henderson and Margaret Sider examined that question in their 2019 paper Build or Renovate? They began with the financing problem. Low-carbon projects could require high upfront investment, long repayment periods and technical knowledge that local lenders lacked. Public capital could prepare projects, absorb particular risks and bring private investors into a market. The institutional choice came after that diagnosis.
Their test covered three sets of facts: the economy and its viable project pipeline; the state's existing programs and institutions; and the development bank's mandate, portfolio, staff, governance, financial health and reputation.
| Test | What must be examined | Why it changes the choice |
|---|---|---|
| The economy | Industrial structure, financial depth, priorities and viable projects | Establishes whether a durable financing problem exists |
| The state | Programs, incentives, agencies and gaps between mandates | Shows what machinery exists and where delivery is fragmented |
| The incumbent bank | Governance, portfolio, staff, financial health and reputation | Determines whether renovation is credible or merely convenient |
A sound public financier already has capital, clients, systems and market access. Giving it a specialist mandate can be faster and cheaper than creating another institution. A weak bank with politically directed lending, poor risk management or a conflicting portfolio can be a bad foundation. A new institution can then be credible despite its cost.
Smallridge and her co-authors also identified risks on the build side. A new bank needs capital, staff, governance and a sufficiently large pipeline. A narrow portfolio is less diversified. A mandate can exceed the resources supporting it. The OECD's wider green-bank research added a crucial qualification: market-mobilization tools assume that limited public intervention can bring private investors into a transaction. Where projects, sponsors and markets are too weak, grants and capacity building may need to come first.
The framework is useful for AI because it resists institutional fashion. It also needs to be widened. Build or Renovate? excluded ordinary agencies, funds, commercial banks and hybrid structures. AI programs routinely combine all four.
AI passes the financing test unevenly
The World Development Report 2026 describes three national pathways: adopting existing AI, adapting it to local needs and advancing the frontier. For most developing countries, building frontier models is not a realistic near-term objective. The urgent work is to establish the foundations that let firms and public services use the technology productively: electricity, connectivity, compute, data, skills and a workable business environment.
Compute is the most visible capital requirement. Global data-centre investment reached about $500 billion in 2024, according to the International Energy Agency, while emerging and developing economies outside China held less than 10 percent of global capacity. Yet a national data centre is not automatically a development project. It needs users, power, connectivity and a revenue model. Smaller economies may receive better value from shared regional facilities or contracted cloud capacity.
Business adoption creates another financing gap. The OECD's review of AI adoption by smaller firms finds that connectivity, data, compute, skills and finance operate together. A company may have to clean its records, connect old systems, redesign a process and train staff before an AI tool produces a return. Much of that investment is intangible and difficult to pledge as collateral.
Infrastructure
Grid, fibre and data-centre shells can support long-lived cash flows.
Compute equipment
Accelerators depreciate quickly and carry residual-value risk.
Young AI companies
Uncertain revenue and intangible assets make ordinary collateral lending a poor fit.
Public goods
Datasets, standards, evaluation and basic research need no fictional repayment stream.
This is the line an AI-finance institution has to hold. It can finance productive capacity and firm adjustment. It cannot become a substitute for an AI ministry, tax reform, social protection or sound regulation.
Four architectures are already visible
Governments are not converging on a national AI development bank. Their current programs show at least four workable architectures.
France
An existing public financier adds AI loans, guarantees, equity, diagnostics, training and advice to an established national platform.
Existing reach and a broad instrument set make another bank difficult to justify.
France is renovating. Bpifrance already provides loans, guarantees, equity, innovation support and advice. It plans to direct €10 billion to the AI value chain by 2030. Its existing firm network and range of instruments make another bank difficult to justify.
India has built a mission. The IndiaAI Mission, approved in 2024 with an outlay of ₹103.72 billion, covers compute, models, datasets, applications, skills, startup finance and trusted AI. Much of its work requires platforms, grants and coordination rather than lending.
Brazil is assembling a mixed platform. Its 2024–28 AI plan proposes R$23.03 billion across infrastructure, training, public services, innovation and governance. Finep, BNDES and other agencies divide the work among credit, grants, equity and research support. In April 2026, BNDESPAR and Finep opened a call for a private manager of a dedicated AI startup fund.
The European Union is testing a regional route. Its proposed €20 billion AI Gigafactories initiative combines the European Commission, the European Investment Bank, member-state budgets and private-led projects through EuroHPC. The model addresses a scale problem that many countries cannot solve efficiently on their own.
These programs are too new to establish which model performs best. They demonstrate that the choice is wider than a new bank or an old one.
A build test for AI finance
A government considering an AI-focused public finance institution should be able to answer six questions before drafting legislation.
Does the evidence support a new institution?
This is an editorial diagnostic, not a substitute for institutional due diligence. “Yes” means the condition for building is present.
The institutional case is still open
Answer all six questions to see which architecture the evidence supports.
What national exposure is it responding to? The diagnosis should identify the industries, fiscal revenues, regions and worker groups at risk, along with the capabilities that could produce new income. A generic ambition to “lead in AI” is not a mandate.
What position in the AI economy is it financing? Broad adoption, local adaptation, data-centre development and frontier research require different systems. The country needs a plausible position grounded in its market, skills, energy system and trading relationships.
Is there a durable pipeline of financeable projects? Announced spending is not a pipeline. Projects need sponsors, users, cash flows, infrastructure and credible demand. Public goods and social spending should be separated from assets that can repay capital.
Which failures call for financial instruments? Guarantees can address collateral or perceived risk. Equity can support young firms. Leasing can manage hardware risk. Grants can prepare projects or purchase public goods. Regulation and worker protection require other institutions.
Can existing institutions carry the mandate credibly? The assessment should cover governance, technical hiring, risk management, equity and guarantee capacity, reach among firms, impact measurement and conflicts with the existing portfolio. Legal permission alone is a poor test.
Would a new institution improve delivery enough to justify its fixed cost? The case is strongest when specialist transactions repeatedly fall between mandates, programs are fragmented, a durable pipeline exists and combined technical and financial teams lower the cost of preparing deals. The institution also needs a method for proving additionality and withdrawing when private finance takes over.
A country that passes these tests may have a case for building. The strongest candidates are likely to be large national or regional markets with significant exposure, a defined development pathway, a durable investable pipeline, gaps in risk-bearing or long-term capital and no institution able to hold the mandate without serious conflict.
Many countries will not pass. Their transition still needs financing. A ring-fenced window in a development bank, a specialist fund, an innovation mission or a regional facility may carry less institutional theatre and deliver more useful capacity.
A country can lose work, revenue and technological position without hosting a frontier AI company. Waiting for the labour data to become conclusive may mean waiting until foreign clients, tax flows and investment patterns have moved. AI warrants the same institutional seriousness that climate finance received—and a stricter test of what, exactly, a bank would finance.
Sources & notes
Exposure estimates classify occupations and tasks; they are not forecasts of job losses. Economic projections use different concepts and scenarios and should not be compared as point estimates. Policy commitments and program figures were reviewed through August 26, 2026. The screening test is Doyen Collective's adaptation of Smallridge et al.