Artificial Intelligence and the Next Development Frontier: Productivity, Capability, and the Institutions That Determine the Gains
I. Introduction: Artificial Intelligence as a Development Technology
Artificial intelligence is commonly discussed as a technological frontier. Its economic importance, however, may ultimately depend less on the sophistication of the underlying models than on the extent to which those models alter the availability, cost, and distribution of useful capabilities across an economy.
Every economy operates under constraints. Capital is scarce. Skilled labor is unevenly distributed. Management quality varies widely across firms. Public institutions often lack sufficient personnel and information. Specialized knowledge that is routine in a major urban center may be effectively unavailable in a rural community. These constraints are particularly consequential in developing economies, where shortages of doctors, teachers, engineers, administrators, financial professionals, and experienced managers can limit the returns to otherwise productive investments.
Artificial intelligence introduces a potentially important change to this structure. A growing range of cognitive tasks can now be assisted by systems capable of interpreting information, recognizing patterns, producing text and code, making forecasts, retrieving knowledge, and supporting decisions. As these systems become cheaper and easier to deploy, some forms of expertise that historically depended on the physical presence of highly trained individuals may become partially replicable through software.
The distinction matters. The central economic question is not whether machines will eventually reproduce human intelligence in some general sense. Nor is it whether every worker will begin using the most advanced available model. The relevant question for development is narrower and more practical: what happens when useful intelligence becomes materially cheaper to reproduce and distribute?
The answer is unlikely to be uniform.
Artificial intelligence may improve the productivity of workers who already possess substantial skills, while also allowing less experienced workers to perform more complex tasks. It may enable firms to operate more efficiently, but it may disproportionately benefit companies with superior data, management, capital, and organizational processes. It may expand public-sector capacity, but it may also amplify administrative mistakes when underlying institutions are weak. It may reduce some barriers between advanced and developing economies while creating new dependencies in computing infrastructure, models, and digital platforms.
For these reasons, AI should not be understood as a shortcut around development. It is better understood as a new production input whose value depends heavily on the system into which it is introduced.
That system includes electricity and connectivity, but also education, organizational capital, competitive markets, usable data, financial depth, administrative competence, and public trust. Countries possessing these complementary conditions are more likely to translate rapidly improving AI capabilities into sustained productivity gains. Countries without them may gain access to the same technologies without realizing comparable economic benefits.
Artificial intelligence may therefore change the production function of development without eliminating the underlying importance of development itself.
II. A General-Purpose Technology Moving at Unusual Speed
Major technological transformations rarely affect productivity immediately.
The steam engine, electrification, automobiles, telecommunications, computers, and the internet all required complementary investment before their full economic effects became visible. Factories had to be reorganized. Infrastructure had to be built. Workers needed new skills. Companies had to redesign production processes. New firms had to emerge while older institutions adjusted.
Artificial intelligence shares several characteristics with these general-purpose technologies. It can be applied across industries. Its capabilities continue to improve. Innovation in one part of the technology creates applications elsewhere. And productive use often requires changes extending well beyond the technology itself.
There is, however, an important difference: the underlying capability is spreading unusually quickly.
The historical diffusion of previous general-purpose technologies across income groups occurred over decades. Estimates suggest that it took roughly 80 years for steam technology to reach lower-income countries, around 40 years for electricity, and roughly 20 years for the internet. By contrast, middle-income economies accounted for approximately half of global ChatGPT traffic within six months of its introduction.
Digital distribution changes the geography of technological diffusion. A model developed in one country can become accessible elsewhere almost immediately. A small business does not have to construct a data center to use an AI application. A doctor does not need to understand model architecture to obtain assistance interpreting clinical information. A programmer can use code-generation tools without owning the infrastructure on which they run.
But rapid access should not be confused with rapid economic transformation.
There is a difference between technological diffusion and economic diffusion.
The first occurs when a technology becomes available. The second occurs when firms and institutions reorganize around it sufficiently to alter output, investment, employment, productivity, and income.
The lag between the two can be substantial.
A company may provide employees with generative AI while leaving workflows unchanged. Workers may use AI primarily for peripheral tasks rather than core production. Public agencies may conduct dozens of pilot programs without integrating successful tools into routine service delivery. Schools may gain access to educational technology without improving learning outcomes. Governments may announce national AI strategies while basic electricity and connectivity remain inadequate.
The economic consequences of AI will therefore depend not only on how rapidly models improve, but also on the speed at which complementary systems adjust.
This distinction argues against both extreme optimism and extreme pessimism. AI need not produce an immediate productivity boom merely because adoption is rapid. Nor does slow measured productivity growth imply that the technology is economically unimportant. General-purpose technologies frequently require substantial institutional and organizational adjustment before their macroeconomic effects become visible.
What is unusual today is that many countries are confronting this adjustment simultaneously.
III. The Economics of Scarce Expertise
One of the most persistent constraints on development is the scarcity of expertise.
The problem is not simply the average level of education. It is the limited supply of specific capabilities at particular places and times.
A rural clinic may have nurses but no specialist physician. A small manufacturer may have workers and equipment but little expertise in inventory optimization or quality control. A local government may collect administrative data without having enough analysts to use it effectively. Farmers may face increasingly volatile weather conditions without reliable access to agronomic advice. Small firms may operate without professional accounting, legal, marketing, or financial expertise because the fixed cost of obtaining such services is too high.
Traditionally, expanding these capabilities has required expanding the supply of trained people. That remains essential. But human-capital formation is slow. Physicians, engineers, teachers, lawyers, and experienced managers take years to train. Migration can redistribute expertise geographically, but often toward rather than away from already prosperous regions.
AI potentially changes this constraint by making parts of expert capability reproducible at very low marginal cost.
This should not be confused with replacing experts.
In many economically valuable applications, the more relevant mechanism is expertise amplification. A physician assisted by diagnostic software may evaluate more patients. A teacher supported by an AI system may produce differentiated exercises for a larger class. An agricultural extension officer may combine local observations with machine-generated forecasts to serve a wider geographic area. A junior employee may perform tasks that previously required considerably more experience.
The economic importance of such systems lies partly in the difference between fixed and marginal costs.
Developing a capable model may be extraordinarily expensive. Once developed, however, providing an additional user with access to some of its capabilities may cost comparatively little. This creates the possibility of separating the production of certain forms of expertise from their distribution.
In economic terms, AI can reduce the marginal cost of access to cognitive capability.
That possibility is especially important where human expertise is scarce rather than where it is already abundant.
Consider a high-income city with many specialists. An AI diagnostic assistant may improve convenience or productivity. In a region where the relevant specialist is effectively absent, the same capability may alter access to a service entirely. The welfare effect can therefore be nonlinear: identical technology can have greater marginal value where the baseline supply of expertise is lower.
The same principle applies to firms.
Large companies have long been able to employ accountants, attorneys, software developers, analysts, designers, translators, and consultants. Small firms cannot always justify these fixed costs. AI services can partially convert professional capabilities previously purchased through expensive labor into variable-cost digital inputs.
This does not remove the need for professional judgment. It does, however, change the boundary of what can be economically provided.
The development implications could be substantial because many persistent productivity gaps are not caused by the absence of technologies in a narrow engineering sense. They are caused by the inability to deploy knowledge consistently across millions of firms, farms, schools, clinics, and public agencies.
AI is potentially powerful because knowledge can travel farther than people.
IV. Access Is Not Productivity
The availability of intelligence does not guarantee its productive use.
This is one of the most important qualifications to the optimistic case for AI-driven development.
Imagine two firms receiving access to exactly the same model. One has reliable digital records, experienced managers, standardized workflows, skilled employees, cybersecurity controls, and sufficient capital to redesign its operations. The other relies on paper records, informal processes, inconsistent data, weak management, and unstable internet access.
The technical capability available to both firms may be identical. The economic return will not be.
AI is therefore strongly complementary with other forms of capital.
Some complements are digital: computing infrastructure, connectivity, cloud services, databases, cybersecurity, and interoperable software.
Others are decidedly analog: electricity, literacy, numeracy, management quality, organizational discipline, financial access, legal institutions, competition, and basic administrative competence.
Human capital remains especially important.
AI systems can reduce the expertise required to perform certain tasks, but users still need enough knowledge to formulate appropriate questions, interpret outputs, identify errors, and decide when automated recommendations should not be trusted. In high-stakes settings, these capabilities can be more important rather than less important as AI becomes more powerful.
This creates an apparent paradox.
The more capable AI becomes, the greater the value of basic complementary capabilities may become.
A worker with domain knowledge can use AI to extend that knowledge. A worker without sufficient foundational skills may have difficulty evaluating plausible but incorrect outputs. A well-managed firm can redesign a workflow around automation. A badly managed firm may add another software subscription without changing production. A capable bureaucracy can use AI to improve targeting or administration. A weak bureaucracy may automate a flawed process and reproduce its errors at larger scale.
The economic return to AI is thus conditional on the quality of the surrounding system.
This has major implications for cross-country convergence.
If advanced economies combine AI with better infrastructure, deeper capital markets, stronger firms, larger pools of skilled workers, and more capable institutions, they may capture larger productivity gains even though the underlying models are globally available.
Technology diffusion alone does not guarantee income convergence.
This pattern is familiar from previous technological revolutions. Computers became widely available long before all firms obtained equivalent productivity gains from them. Research on information technology repeatedly found that organizational capital and management practice influenced the returns to digital investment.
AI may intensify the same mechanism because it interacts directly with decision-making and knowledge work.
The relevant unit of adoption is consequently not the software license.
It is the workflow.
V. Labor Markets: Augmentation Before Displacement
Public discussion of artificial intelligence frequently begins with a binary question: which jobs will disappear?
That framing is incomplete.
Jobs are collections of tasks. Technologies rarely automate every task within an occupation at the same time. They alter the relative cost of performing some activities, increase the value of others, and create incentives to reorganize production.
The immediate labor-market effects of generative AI are also likely to differ materially across income groups.
Employment in higher-income economies is relatively concentrated in cognitive, administrative, professional, and service occupations whose tasks can often be performed digitally. Developing economies employ a larger share of workers in agriculture, construction, manufacturing, transportation, retail, and other activities requiring substantial physical interaction.
As a result, current exposure to generative-AI automation is considerably lower in developing economies. Recent estimates place the share of existing jobs exposed to meaningful automation risk at approximately 4.5 percent in low- and middle-income economies, compared with 14.2 percent in high-income economies. Yet the estimated share of jobs that could experience meaningful productivity enhancement is much closer: approximately 16.2 percent in developing economies against 18.7 percent in high-income countries.
These figures suggest an important asymmetry.
For many developing economies, the near-term economic opportunity from AI may be larger through augmentation than through labor substitution.
That distinction matters for policy.
If AI primarily complements workers, the relevant objective is not to protect existing tasks from technological change. It is to increase workers' ability to use new tools productively and to facilitate movement toward activities whose value rises when cognitive capabilities become cheaper.
This does not mean displacement risks are negligible.
Some developing economies have built export industries around digitally deliverable services precisely because skilled labor is cheaper than in advanced economies. Business-process outsourcing, customer support, basic coding, translation, content production, and administrative services may be among the earliest activities exposed to generative automation.
AI could therefore weaken some established development strategies even while creating opportunities elsewhere.
The effect will depend on whether firms compete primarily by providing inexpensive routine cognitive labor or by combining human judgment, customer knowledge, specialized expertise, and AI tools into higher-value services.
Labor-market adjustment will also be uneven within countries. Workers with stronger digital skills may gain earlier. Large urban firms may adopt faster than informal enterprises. Younger workers may adapt differently from older workers. Some occupations may experience higher productivity but lower employment if output demand does not expand sufficiently.
For these reasons, measuring "AI exposure" is only a starting point.
The central question is how technology changes the marginal productivity of different workers and how markets respond to those changes.
VI. Firms, Competition, and the Distribution of Productivity Gains
Aggregate productivity is ultimately generated within organizations.
This makes the firm a critical unit for understanding whether AI becomes a broadly shared economic technology or remains concentrated among a relatively small number of leading companies.
The optimistic mechanism is straightforward.
AI can reduce the cost of performing administrative, analytical, marketing, programming, translation, customer-service, research, and design tasks. It can help small firms obtain capabilities that previously required larger teams or external professional services. By lowering certain fixed costs, AI may make entrepreneurship easier and allow younger firms to compete in markets that once favored organizational scale.
The opposing mechanism is equally plausible.
Large firms possess more proprietary data, stronger management systems, better access to capital, larger technology budgets, and more specialized employees. They can integrate AI across entire production systems rather than use isolated consumer applications. They are also better positioned to negotiate access to computing infrastructure, recruit scarce technical talent, and absorb the fixed costs associated with cybersecurity, governance, and model customization.
AI can therefore lower some barriers to entry while raising the return to scale in other dimensions.
The distribution of productivity gains will depend heavily on market structure.
Suppose a small number of frontier firms achieve large productivity improvements while thousands of smaller firms change little. Measured aggregate productivity may rise modestly, but economic concentration can increase sharply. Labor and capital may migrate toward leading firms, yet inefficient firms may survive for extended periods where competition is weak, finance is distorted, or exit is politically difficult.
By contrast, competitive product markets can accelerate technological diffusion. Firms facing strong competitive pressure have greater incentives to adopt productivity-enhancing tools. New entrants can build AI-intensive organizations without inheriting legacy systems. Workers move toward more productive firms. Capital is reallocated toward scalable business models.
This is why an effective AI strategy cannot be separated from the broader business environment.
Policies that improve access to finance, simplify firm entry, permit unsuccessful firms to exit, protect competition, reduce regulatory uncertainty, and facilitate experimentation may matter as much as direct subsidies for AI adoption.
There is also a management problem.
Installing software is easy. Reorganizing a company is difficult.
The largest productivity gains may emerge when firms redesign workflows around the new technology rather than simply inserting AI into existing processes. That can require changing responsibilities, approval structures, employee incentives, customer interfaces, data architecture, and even the boundaries of the firm itself.
Such reorganization takes time.
This helps explain why an economy can experience spectacular improvements in AI capabilities without immediately recording an equally spectacular acceleration in measured productivity.
The technology moves first. Organizations follow.
VII. Concentration in the AI Economy
Artificial intelligence is digitally distributed but industrially concentrated.
That combination distinguishes the current technological transition.
At the application layer, entry barriers can appear low. Developers can build products using externally provided models and cloud infrastructure. Individuals can access sophisticated systems through ordinary devices. Innovation can occur far from the companies that trained the underlying models.
Beneath this accessible surface, however, the production structure is considerably more concentrated.
Advanced semiconductor manufacturing, high-performance computing, hyperscale cloud infrastructure, frontier-model development, and certain critical software ecosystems require enormous capital expenditure, technical expertise, energy availability, and scale.
A relatively small number of firms and economies therefore control important bottlenecks in the AI value chain.
For developing economies, this concentration creates two opposing effects.
The first is dependency.
Countries that rely heavily on external providers may face changes in prices, access conditions, model policies, data requirements, or geopolitical restrictions that they cannot control. Public agencies could become dependent on proprietary systems. Domestic firms may build products whose economics depend on a small number of upstream suppliers. Sensitive data may interact with systems governed under foreign jurisdictions.
The second effect is accessibility.
Precisely because expensive infrastructure is concentrated, every country does not need to reproduce it.
The global cloud model allows businesses to rent computing capacity rather than own data centers. Application developers can use pretrained models rather than train frontier systems from scratch. Governments can adapt existing technologies instead of committing scarce fiscal resources to projects with questionable commercial viability.
Concentration therefore generates dependency risk but also creates economies of scale that make advanced capability affordable to countries that could never independently finance the full technology stack.
The appropriate policy response is not self-evident.
Treating every layer of the AI value chain as strategically indispensable would lead many countries toward extraordinarily expensive industrial policies. Yet ignoring concentration entirely could create forms of technological dependence that become costly later.
The more useful objective is resilience.
Countries should determine which capabilities require domestic control, which require diversified access, and which can safely be purchased internationally.
This is a portfolio problem rather than a binary choice between sovereignty and dependence.
VIII. Adaptation May Matter More Than Sovereignty
National AI strategies increasingly use the language of sovereignty.
The motivation is understandable. Technology has become deeply intertwined with national security, economic competitiveness, data governance, and geopolitical power. Governments worry that dependence on foreign chips, cloud platforms, or models could constrain future policy choices.
But sovereignty is an expensive concept when applied indiscriminately.
Building frontier AI systems requires scarce technical talent, advanced semiconductors, enormous data-center investment, significant electricity supply, sophisticated capital markets, and continuous spending to remain competitive as the technological frontier moves.
For a limited number of economies, these investments may be commercially and strategically defensible.
For many others, attempting to reproduce the entire AI stack could consume scarce public resources while generating little durable comparative advantage.
The relevant distinction is between strategic autonomy and technological autarky.
Autonomy requires retaining meaningful choice.
A country may want multiple cloud providers, portable data, interoperable systems, cybersecurity capacity, domestic expertise capable of evaluating external technologies, and the legal ability to switch vendors.
None of those objectives necessarily requires constructing a national frontier model.
The highest economic returns may instead come from adaptation.
A general model trained primarily on data from advanced economies may have remarkable capabilities while remaining poorly suited to a particular legal system, language, agricultural environment, curriculum, clinical workflow, or administrative process.
Local economic value emerges when generalized technical capability is connected with specific information.
This can involve fine-tuning models, improving retrieval systems, developing local-language datasets, combining AI with sector-specific knowledge, designing interfaces for different levels of literacy, or building applications that function under limited bandwidth.
Such investments are less glamorous than constructing enormous data centers, but they may produce higher social returns.
This suggests a broader principle for technological development.
Countries do not become productive by owning every technology they use. They become productive by developing the capability to select, absorb, adapt, combine, and improve technologies created throughout the world.
The history of industrialization contains many examples of economies that advanced through technological absorption before becoming important sources of frontier innovation themselves.
AI should not be expected to follow a fundamentally different economic logic.
IX. Localization as an Economic Capability
AI is unusually sensitive to context.
A machine can execute the same mathematical operations anywhere, but useful decisions depend on local information.
Medical recommendations depend on disease prevalence, treatment protocols, available medicines, clinical infrastructure, and population characteristics. Agricultural advice depends on soil, weather, crop varieties, market conditions, and local farming practices. Legal assistance depends on jurisdiction. Educational tools depend on language, curriculum, teaching methods, and student ability.
Even communication itself can create barriers. Many populations interact more naturally through speech than through written prompts. Some languages have abundant digital training data; others do not. Some users possess smartphones and reliable broadband; others rely on inexpensive mobile devices and unstable networks.
The economic value of AI therefore depends partly on whether systems are designed around actual users rather than idealized users.
This creates an important innovation frontier that is separate from the race to build the world's most capable general model.
For many markets, the economically relevant technologies may be smaller, cheaper, specialized systems optimized for specific tasks.
These systems may run with limited computing requirements. They may combine voice interfaces with local-language knowledge. They may operate on devices rather than exclusively in large data centers. They may connect a general-purpose model to verified domain-specific databases. They may sacrifice broad capability in exchange for reliability, affordability, privacy, or ease of use.
This category of "small AI" has particular relevance for developing economies because it changes the cost structure of deployment. AI can potentially reach users through familiar technologies rather than requiring every household, school, clinic, or small business to adopt frontier computing infrastructure.
Localization should consequently be treated as an economic capability in its own right.
Countries need firms and institutions capable of understanding both the imported technology and the local problem.
That combination is not automatic.
It requires engineers who understand domestic industries, domain experts willing to work with technology teams, accessible data, financing for experimentation, and customers capable of distinguishing effective solutions from impressive demonstrations.
The competitive advantage of many developing economies may therefore lie not in training the largest model, but in developing superior knowledge about where and how models should be used.
Frontier intelligence may become increasingly global.
Context remains local.
X. AI and the Capacity of the State
The productivity discussion should not end with firms.
In many economies, the quality of the state is itself an important constraint on development.
Governments must collect taxes, maintain registries, administer benefits, forecast weather, inspect infrastructure, deliver education and health services, process legal cases, regulate markets, respond to disasters, and allocate public resources. These activities are information-intensive.
Administrative capacity is also unevenly distributed.
Some governments possess sophisticated digital systems and large pools of technical expertise. Others operate with fragmented databases, paper records, staffing shortages, and limited analytical capability.
Artificial intelligence could improve state capacity through two broad channels.
The first is back-office productivity.
Machine learning can help classify records, detect anomalies, prioritize inspections, forecast demand, identify tax irregularities, translate documents, summarize administrative material, and assist case management. These uses can improve the productivity of existing public employees without fundamentally changing the interface between government and citizens.
The second is front-line capability.
AI systems could assist teachers, health workers, agricultural extension officers, legal-aid providers, and other public servants who interact directly with citizens. In locations where specialists are scarce, the productivity gains could be particularly large.
Yet public-sector adoption raises a stricter standard than many private applications.
A consumer can stop using an unreliable chatbot. A citizen cannot necessarily opt out of a tax authority, welfare agency, court, or immigration system.
Errors can therefore have coercive consequences.
An AI system used to draft advertising copy may tolerate occasional mistakes. A system influencing medical treatment, legal decisions, benefit eligibility, or law enforcement requires stronger evidence, auditability, accountability, and human oversight.
Governments need institutional capability not only to purchase AI but to evaluate it.
This is likely to become a significant bottleneck.
Public sectors have a long history of technology pilots that fail to scale. AI may worsen the problem because demonstrations can appear persuasive even when evidence of actual social benefit remains weak.
The challenge is to develop procurement systems that reward measurable outcomes rather than technological novelty.
Governments must be able to test systems, compare alternatives, assess costs, monitor performance after deployment, protect sensitive information, and discontinue tools that do not work.
AI can amplify state capacity.
It can also amplify state incapacity.
The difference lies in governance.
XI. The Foundations Beneath the Algorithm
The AI era may reinforce some of the oldest lessons of development economics.
Advanced software still requires electricity.
Cloud services still require connectivity.
Digital public services still require citizens capable of accessing them.
Automated decisions still require reliable information.
Innovative firms still require capital and competitive markets.
Sophisticated technology does not repeal basic constraints.
In parts of Sub-Saharan Africa, nearly one-third of rural schools still lack reliable electricity and more than two-thirds lack dependable internet access. Under such conditions, discussion of frontier AI can become detached from the infrastructure needed to deliver even ordinary digital services.
This does not imply that countries must complete one stage of development before beginning another.
Technology can permit leapfrogging. Mobile networks allowed many economies to expand telecommunications without reproducing the fixed-line infrastructure of richer countries. Mobile money developed rapidly in environments where traditional banking penetration was limited.
AI may produce similar discontinuities.
But leapfrogging technologies still require something to land on.
A voice-based agricultural assistant can reduce literacy barriers, but it still requires communications infrastructure. A diagnostic system can extend scarce medical expertise, but someone must collect data and deliver treatment. AI tutoring may complement teachers, but students still need basic literacy and sufficient access to devices.
The economic implication is that AI should increase rather than reduce the expected return to foundational investment.
Reliable power becomes more valuable when electricity supports not only traditional production but also digital services and computing. Broadband becomes more productive when it provides access to increasingly sophisticated capabilities. Education becomes more important when skilled users can leverage AI to multiply their output. Administrative data becomes more valuable when algorithms can convert it into operational insights.
There is therefore a danger in treating national AI strategy as a separate policy domain.
The most effective AI policy may often look surprisingly conventional: better schools, more reliable electricity, competitive telecommunications, stronger firms, deeper financial markets, interoperable public data systems, and competent institutions.
The algorithm sits at the top of a much larger economic structure.
When the foundations are weak, technological capability remains partially stranded.
XII. Regulation, Trust, and Institutional Discipline
AI policy must address a genuine tension.
Regulation imposed too early or designed too broadly can raise entry costs, protect incumbents, slow experimentation, and create compliance burdens that governments themselves lack the capacity to administer.
But an absence of governance can generate privacy violations, cybersecurity vulnerabilities, discrimination, fraud, unsafe systems, market concentration, political manipulation, and declining public trust.
The appropriate objective is therefore neither maximal regulation nor maximal permissiveness.
It is institutional proportionality.
Different applications create different risks.
A recommendation system for entertainment content does not warrant the same oversight as an AI system used in medical diagnosis. Software that assists an employee drafting an internal memo should not be regulated identically to a model influencing criminal justice or social-benefit eligibility.
Risk-sensitive governance allows scarce regulatory capacity to be concentrated where errors are most consequential.
For countries with limited administrative resources, there is also a sequencing problem.
Attempting to build a comprehensive AI regulatory architecture before regulators understand the technology, before domestic markets have developed, and before evidence about specific harms is available can produce rules that are both costly and ineffective.
A more pragmatic approach begins with technical and industry standards, applies existing laws where relevant, develops targeted rules around demonstrated risks, and strengthens enforcement capacity over time. International coordination is particularly valuable because inconsistent national rules can fragment digital markets without meaningfully reducing systemic risks.
Trust is an economic variable in this process.
People will be less willing to provide data or adopt AI-enabled public services if they expect surveillance, arbitrary decisions, or weak privacy protection. Firms will invest less when liability rules are unpredictable. Governments will have difficulty scaling successful systems if public resistance becomes generalized.
At the same time, trustworthy governance should not become a rhetorical substitute for measurable performance.
The legitimacy of AI systems will ultimately depend partly on whether they produce better outcomes.
If technology helps schools teach more effectively, clinics reach more patients, governments process services more quickly, and firms offer better products at lower cost, adoption becomes easier to sustain.
Institutional discipline therefore requires both protection against harm and evidence of benefit.
XIII. Adopt, Adapt, Advance
Different economies should pursue different positions on the AI frontier.
A useful strategy can be understood through three increasingly demanding stages: adoption, adaptation, and advancement.
Adoption begins with existing tools.
This is the lowest-cost path and is likely to generate the fastest returns. Firms can introduce AI into routine workflows. Workers can use it to complement existing expertise. Governments can test applications in administration and public services.
The economic case for adoption is strongest where proven technologies can solve clearly defined problems without requiring large domestic investments in underlying infrastructure.
Adaptation goes further.
Generic AI is modified around local language, data, regulation, industrial structure, user behavior, and institutional conditions.
For many developing economies, adaptation may offer the highest return on scarce resources.
It requires more capability than simple adoption, but far less than building the technological frontier from first principles. It can also create domestic firms and expertise with genuine comparative advantage because successful adaptation depends on local knowledge that global technology providers may not possess.
Advancement involves contributing directly to the frontier: developing advanced models, semiconductor capabilities, large-scale computing infrastructure, novel architectures, or other foundational technologies.
The costs are substantially higher.
Advancement can make sense where countries possess the necessary scientific base, capital, infrastructure, energy supply, industrial capabilities, and market scale. Elsewhere, prestige-driven efforts to reproduce technologies already available internationally may generate poor returns.
These stages should not be treated as a hierarchy of national status.
A country does not fail because it chooses not to develop a frontier model.
The objective of economic policy is not technological symbolism. It is productivity, resilience, income growth, and welfare.
Nor must every country move mechanically from one stage to the next.
Some may remain exceptionally successful adapters. Others may advance in specialized parts of the AI stack while relying on international markets elsewhere.
The broader principle is straightforward:
The sophistication of AI policy should follow the sophistication of economic capability.
Public investment should expand where there is evidence that domestic capabilities justify it, rather than assuming that every emerging technology must become an indigenous national industry.
Sequencing matters because resources are scarce.
A dollar devoted to subsidizing a frontier data center is a dollar that cannot simultaneously finance electricity networks, education, research universities, digital infrastructure, health systems, or other public priorities.
AI strategy is therefore ultimately an exercise in capital allocation.
XIV. The Risk of a New Productivity Divide
The rapid global diffusion of AI has encouraged the view that developing economies may be able to narrow technological gaps more quickly than during earlier industrial transitions.
That possibility is real.
It is not guaranteed.
If AI capabilities become globally accessible while the complementary assets required to use them remain highly unequal, technological convergence could coexist with productivity divergence.
Consider two economies with access to the same advanced model.
One combines it with reliable electricity, universal broadband, high educational attainment, deep capital markets, digitized firms, high-quality management, large datasets, world-class universities, sophisticated cloud infrastructure, and effective public institutions.
The other combines the model with intermittent electricity, expensive connectivity, limited technical skills, weak credit markets, fragmented data, small informal firms, and low administrative capacity.
The model may narrow the gap in available intelligence.
The economic gap could nevertheless widen.
This would represent a new form of digital divide.
The earlier digital divide was often framed primarily around access: who had a computer, a mobile connection, or internet service.
The emerging divide is more demanding.
It concerns the capacity to translate access into productivity.
An AI productivity divide could emerge between countries, regions, firms, and individuals that can reorganize around inexpensive intelligence and those that cannot.
The distributional consequences could extend through trade.
Countries whose export advantage rests heavily on routine cognitive labor may face pressure if AI reduces the cost of performing those tasks elsewhere. Economies capable of combining AI with manufacturing, logistics, advanced services, research, or specialized domain knowledge may benefit disproportionately.
Capital ownership also matters.
If AI raises returns to intangible assets, proprietary data, computing infrastructure, and highly scalable firms, income may shift toward owners of capital even where aggregate productivity improves.
Policy should therefore not evaluate AI solely by asking whether it raises GDP.
The more relevant questions include who captures the gains, whether new firms can enter, whether workers can move toward more productive activities, whether regions outside major cities participate, and whether public services improve for populations that previously lacked access.
Technological progress can increase the size of the economic pie while changing its distribution.
Both effects matter.
XV. Conclusion: Development in an Age of Abundant Intelligence
Modern economic development has always been shaped by scarcity.
Capital is scarce. Energy is scarce. Skilled workers are scarce. Management capability is scarce. Reliable information is scarce. Institutional capacity is scarce.
Artificial intelligence does not eliminate these constraints.
It may, however, change the economics of one of them.
Useful cognitive capability is becoming cheaper to reproduce.
A farmer may obtain advice that previously required access to an agronomist. A small company may obtain analytical support previously affordable only to a large corporation. A physician may receive assistance interpreting information outside a narrow specialty. A teacher may produce individualized materials at much lower cost. A government employee may analyze records that previously remained unused.
None of these applications requires the machine to become universally intelligent.
They require only that AI performs economically valuable tasks at sufficiently low cost and acceptable reliability.
That narrower development is already consequential.
If the marginal cost of certain forms of expertise declines substantially, production can reorganize around a new abundance. Some occupations will change. Some firms will become more productive. Some business models built around scarcity of routine cognitive labor will come under pressure. New services will become economically viable. Human expertise will shift toward tasks where judgment, accountability, relationships, physical interaction, and contextual understanding remain important.
But intelligence will not become productive merely because it is available.
AI must still be combined with electricity, networks, skills, capital, data, organizations, markets, and institutions.
This is why the strongest AI strategies may look less futuristic than expected.
Countries need sufficient infrastructure to connect people and firms. Education systems must provide foundational capabilities that allow workers to use increasingly powerful tools. Competitive markets must allow productive firms to expand. Capital must move toward new opportunities. Public agencies must become capable buyers and users of technology. Regulation must protect trust without freezing experimentation. Domestic businesses must learn how to adapt global technologies to local conditions.
The countries that succeed will not necessarily be those that spend the most on AI.
They will be those that allocate resources well.
For advanced economies, the challenge may increasingly involve managing disruption as highly exposed cognitive occupations change rapidly. For developing economies, the immediate opportunity may be different: using AI to alleviate shortages of expertise that have constrained productivity and public-service delivery for decades.
That opportunity creates a potentially important asymmetry.
Developing economies do not necessarily need to reproduce the enormous fixed investments required to create frontier intelligence. They can increasingly purchase, access, and adapt capabilities generated elsewhere.
Their comparative advantage may emerge in the application layer: identifying high-value problems, combining AI with local information, redesigning services, and building organizations capable of turning inexpensive intelligence into useful output.
This is not a minor role.
Most of the economic value created by a general-purpose technology need not accrue to the companies that invent its foundational components.
Electricity transformed industries far removed from electric utilities. Computing generated enormous value through businesses that never manufactured processors. The internet reshaped retail, finance, media, logistics, and communications through organizations built on infrastructure created elsewhere.
AI may follow a similar pattern.
The frontier matters. But diffusion matters more for aggregate welfare.
The greatest uncertainty is therefore not how powerful AI systems will ultimately become. Technological progress will continue to surprise in both directions, and predictions about specific capabilities remain unusually fragile.
The more durable economic question is whether institutions can absorb the capability already becoming available.
If they can, artificial intelligence could weaken one of the persistent constraints on development: the limited availability of specialized knowledge.
If they cannot, the same technology may reinforce existing advantages, allowing already productive firms, workers, cities, and countries to pull further ahead.
The distinction is critical.
Artificial intelligence is neither a substitute for development nor merely another sector within it. It is becoming an input into development itself.
Its promise lies less in replacing labor than in lowering the cost of accessing expertise; less in eliminating institutions than in increasing the returns to competent ones; and less in making geography irrelevant than in allowing knowledge to travel across geography at unprecedented scale.
The economies that convert that possibility into sustained growth will be those that recognize a simple principle: technological capability and economic capability are not the same thing.
The first can increasingly be imported.
The second must still be built.
Comments