World Bank Small AI Strategy Challenges the Race for Bigger Models

The World Bank small AI strategy is urging developing countries to pursue focused, mobile-ready systems, despite an industry race toward larger and more expensive models. These tools target specific problems, use narrower datasets, and can work with limited connectivity. The approach shifts attention from benchmark leadership to whether AI can deliver useful expertise through devices and networks that people already have.

World Bank chief knowledge officer Paschal Donohoe made the case during the World Bank and International Monetary Fund meetings in Bangkok. He told the Guardian that locally adapted AI could compress development gains that once required centuries into decades. The claim is ambitious, but its underlying conflict is immediate: frontier AI depends on costly infrastructure concentrated in wealthy economies, while small AI aims to work within local constraints.

The strategy does not argue that smaller systems can replace every frontier model. It argues that most developing economies should prioritize adoption and local adaptation before trying to build the most advanced systems themselves. That choice pressures governments, development institutions, and technology vendors to prove that specialized AI can scale beyond isolated pilots.

The World Bank Small AI Push Changes the Development Goal

The World Bank is redefining AI progress around accessible services, not the size of the underlying model.

Small AI refers to affordable systems designed for narrow tasks, local languages, and ordinary devices. A model might analyze a crop image, support medical screening, or deliver advice through a voice call. It does not need to answer every question or compete across broad academic benchmarks.

That narrower scope distinguishes the approach from the general-purpose models developed by companies such as OpenAI, Google, Anthropic, and Meta. Those systems require extensive computing capacity during training. Their most capable services also depend on cloud infrastructure and reliable connectivity during use.

The World Bank argues that developing countries can start elsewhere. Its development roadmap recommends three stages: adopt existing tools, adapt them to local conditions, and advance toward frontier development only when national capabilities justify the expense.

This sequencing is central to the proposal. A government does not need its own general-purpose model before deploying AI-supported agricultural advice. A clinic does not need constant broadband if a specialized screening system can operate locally and synchronize later.

Donohoe described that distinction through AgriConnect, a World Bank agricultural initiative in India. Farmers need soil information that applies to their crops, locations, and languages. The country cannot place an agricultural scientist beside every small farm, but a phone-based system could distribute parts of that expertise.

The model still needs trustworthy agricultural data and local validation. It also needs an interface that farmers can use, which may mean voice rather than text. Small AI therefore describes more than a compact model. It is a delivery strategy organized around a defined user, a constrained environment, and a measurable task.

The World Bank’s case arrives when AI investment remains concentrated around large models and the infrastructure supporting them. That market rewards general capabilities, global platforms, and frequent model upgrades. Development projects usually need something less glamorous: predictable performance on a local problem, under difficult operating conditions.

This is why the small AI interview matters. The World Bank is asking countries to treat localization as the main product, not as a secondary feature added after deployment.

That change also alters how success should be measured. A smaller system does not need to outperform a frontier model on every test. It needs to produce a useful result at an acceptable cost, in the required language, on the available hardware, with safeguards suited to the setting.

The proposal turns the development question from “Can this country build a leading model?” into “Can this country deliver better decisions where expertise is scarce?” That is a more practical goal, but it also demands stronger evidence from the field.

Why Bigger Models Fit Poorly With Local Constraints

The frontier model race assumes infrastructure that much of the developing world does not yet possess.

The World Bank’s 2026 report says AI is spreading faster than earlier general-purpose technologies. The steam engine took about 80 years to reach lower-income countries, electricity took roughly 40 years, and the internet took about 20 years. Middle-income countries generated half of ChatGPT’s global traffic within six months of its launch.

Fast diffusion, however, does not mean equal access. The World Bank says 2.2 billion people remain offline. Its small AI brief also reports that low-income countries account for less than one percent of ChatGPT usage.

That divide becomes sharper when infrastructure is considered. In Sub-Saharan Africa, nearly one-third of rural schools lack reliable electricity. More than two-thirds lack dependable internet access, according to the Bank’s AI development findings.

A cloud service can be technically available while remaining practically unreachable. Unstable power interrupts devices and network equipment. Expensive mobile data limits repeated use. High latency can make interactive tools frustrating. A text-first interface excludes people who cannot read the supported language.

Frontier development creates another barrier. Advanced models require specialized chips, large data centers, extensive training data, and experienced researchers. These assets remain concentrated among a small number of companies and countries.

The World Bank’s earlier AI foundations report found that high-income countries accounted for 87 percent of notable AI models as of July 2025. They also held 86 percent of AI startups and 91 percent of cumulative venture funding for AI startups.

Those countries represented 17 percent of the global population. The imbalance means most developing economies will consume technology built elsewhere unless they find a workable path for local adaptation.

Small AI tries to reduce that dependency without pretending it can eliminate it. A specialized model can be adapted from an existing open or proprietary foundation. It can use fewer computing resources, operate on a phone or laptop, and focus on data relevant to a particular region.

This does not make infrastructure irrelevant. Every AI system still depends on electricity, devices, maintenance, and some form of data pipeline. Training or updating a local model may require cloud computing even when daily use happens offline.

The difference is one of thresholds. Small AI lowers the minimum infrastructure needed to deliver a service. It lets institutions begin with a bounded application while improving connectivity, skills, and computing capacity over time.

The World Bank organizes those requirements around four foundations: connectivity, compute, context, and competency. Connectivity covers electricity and internet access. Compute includes chips, data centers, and cloud services. Context means relevant data, institutions, and languages. Competency includes technical skills and the ability to use systems responsibly.

Weakness in any one area can stop a deployment. A health model with strong accuracy still fails if a clinic cannot maintain its device. A farming assistant cannot help if its recommendations reflect another region’s soil and weather. A local-language interface is insufficient when users do not trust the institution collecting their data.

The pressure therefore falls on more than technology companies. Governments must improve basic infrastructure and procurement. Universities and training programs must develop local expertise. International institutions must fund evaluation, not only pilots. Vendors must support interoperable systems rather than locking public services into one platform.

Local Adaptation Is the Mechanism, Not a Feature

Small AI works only when local knowledge changes the model, interface, and service around it.

The World Bank argues that importing an AI service is not the same as adapting it. A model trained mainly on high-income settings can generate advice that conflicts with local laws, agricultural conditions, medical practices, or social norms. Translation alone cannot resolve those differences.

Language is one visible part of the problem. Models often perform unevenly across languages, particularly when high-quality training data is scarce. Even accurate written output will miss users who rely on spoken communication or basic phones.

The Bank proposes delivery through text messages, voice calls, and lightweight applications. These channels matter because they match existing behavior. A farmer can ask a question by voice. A health worker can use an offline screening tool. A teacher can receive lesson support without maintaining a constant connection to a large cloud model.

This is an information-design challenge as much as a model-design challenge. Institutions must combine technical output with local records, professional judgment, and operational knowledge. Similar knowledge blending principles matter whenever general AI must work alongside specific, trusted sources.

Agriculture offers a clear example. A useful system needs more than a generic understanding of crop science. It may need regional weather, soil conditions, pest patterns, planting calendars, and advice approved by local experts.

Health applications face higher stakes. The World Bank points to handheld tuberculosis screening systems that do not require constant broadband. Such tools can extend limited clinical capacity, but they must be validated against the population and conditions where they operate.

An incorrect agricultural recommendation can reduce a harvest. An incorrect medical recommendation can delay treatment. Narrow scope reduces some risks because developers can define and test the intended task. It does not remove the need for human review or escalation.

Education tools present a similar tradeoff. Lightweight tutors can offer practice and feedback where teachers have limited time. Yet learning gains depend on curriculum alignment, language quality, device access, and whether teachers can monitor the system’s use.

Small AI can complement professionals by handling repetitive analysis or extending their reach. It should not become an excuse to avoid investing in teachers, doctors, agricultural specialists, or public institutions.

The most credible deployments will define exactly where the system assists a person. A crop tool might flag a likely pest and recommend that uncertain cases go to an extension worker. A screening device might prioritize patients for clinical review rather than issuing a final diagnosis.

Local adaptation also changes who controls the resulting service. Governments need access to performance data, audit procedures, and methods for switching providers. Otherwise, a compact model can create the same vendor dependency as a large cloud platform.

Open-source models can lower barriers because developers can inspect and modify more of the system. They also shift responsibilities toward local organizations, which must maintain software, manage security, and verify updates.

Proprietary models may provide better support or stronger general capabilities. They can also expose institutions to changing terms, remote service interruptions, and pricing decisions outside local control. The right choice will vary by application, but procurement should preserve access to local data and operational continuity.

Interoperability is therefore part of the mechanism. The World Bank recommends that national systems work with multiple platforms and providers. That reduces the risk that one vendor becomes inseparable from a public service.

Data governance matters just as much. Local adaptation often requires sensitive health, education, financial, or agricultural records. Institutions need clear rules about consent, retention, reuse, and access. Communities also need a practical way to challenge harmful decisions.

The development value comes from combining a capable model with a trusted delivery system. Without that surrounding structure, small AI is only smaller software.

The Evidence Gap Could Swallow the Promise

The largest uncertainty is not whether small models can run on limited hardware, but whether successful pilots can become dependable public services.

The World Bank presents real applications in agriculture, health, and education. These examples show that specialized AI can operate under constraints. They do not establish that every deployment will produce durable gains at national scale.

Pilots often receive dedicated funding, expert attention, and carefully selected users. Expansion introduces harder conditions. Devices break. Staff members change. Data drifts. Local agencies inherit maintenance costs that were absent from the initial trial.

A model can also perform well in a laboratory and fail with real users. Photos may be blurry. Accents may differ from training samples. Medical equipment may vary between clinics. Farmers may describe symptoms using local terms that the system does not recognize.

This is why evaluation must measure service outcomes, not only model accuracy. A crop assistant should be judged by whether farmers make better decisions and protect yields. A screening tool should be judged by clinical outcomes, missed cases, false alerts, and referral quality.

Governments also need comparisons with simpler alternatives. A new AI tool might appear useful while delivering no more value than a searchable guide, a better call center, or additional training for frontline workers. Limited development budgets make that comparison essential.

The Bank acknowledges that countries need better procurement and evaluation frameworks. Many pilots are already underway, but institutions do not always know which ones deliver results. Scaling weak projects would waste scarce resources and damage public trust.

Bias creates another risk. Local data may be incomplete or reflect unequal access to public services. Training a model on those records can reproduce those patterns. A small model may also have less capacity to handle unusual cases outside its narrow domain.

Privacy risks do not shrink with parameter count. A phone-based application can still collect sensitive health information, location data, voice recordings, or farm records. Offline operation may reduce some exposure, but synchronization and model updates create additional points of risk.

Security also becomes difficult when devices are widely distributed. Older phones may lack current protections. Local applications may remain unpatched. Attackers could manipulate inputs, steal data, or alter recommendations.

There is a political risk as well. Public agencies might use AI to automate eligibility, enforcement, or surveillance without adequate appeal mechanisms. A system designed for efficiency can weaken rights if affected people cannot understand or contest its decisions.

The labor impact requires similar restraint. The World Bank estimates that 4.5 percent of existing jobs in low- and middle-income economies face automation risk from generative AI. The comparable estimate for high-income countries is 14.2 percent.

At the same time, the Bank estimates that AI could meaningfully raise productivity in 16.2 percent of jobs in developing economies. That is close to the 18.7 percent estimate for high-income countries.

These figures support a complementing-workers strategy, but they remain economy-wide estimates. Actual outcomes will depend on sector, occupation, employer behavior, and worker bargaining power. Productivity gains do not automatically become better wages or improved working conditions.

The World Bank’s position is that developing countries should focus on augmenting scarce expertise. That is plausible in sectors where too few professionals serve large populations. It becomes less convincing if institutions use AI primarily to reduce staffing or weaken accountability.

Another uncertainty concerns market concentration. Small AI can be built on top of technology controlled elsewhere. A country may avoid constructing a data center but still depend on foreign chips, cloud services, model weights, or application programming interfaces.

Localization can therefore reduce one form of dependency while preserving another. Governments should examine the entire supply chain, including who controls updates, data access, and technical support.

The World Bank’s claim that AI can compress a century of development into a decade should be read as a statement of potential, not a forecast. Economic development depends on institutions, education, infrastructure, public health, trade, finance, and political stability. Software cannot substitute for those foundations.

Small AI is most credible when it improves a defined service within that larger system. It becomes less credible when presented as a shortcut around basic investment.

Three Signals Will Show Whether Small AI Can Scale

The next test is whether governments can turn specialized models into measured, maintainable services without creating new dependencies.

The first signal is rigorous evidence from scaled deployments. Governments and development institutions should publish results that compare AI-supported services with existing methods. The most useful studies will include error rates, operating costs, user outcomes, and performance across languages and regions.

Positive results from agriculture, health, or education would strengthen the World Bank small AI thesis. Repeated pilot failures or weak comparisons would show that low-cost deployment does not guarantee meaningful impact.

Evidence must also include the users who are easiest to exclude. Systems should be tested with people using older phones, weaker connections, different dialects, and limited digital literacy. Average performance can hide serious access gaps.

The second signal is investment in the four foundations. Small AI reduces infrastructure requirements, but it still needs electricity, connectivity, computing access, local data, and skilled people. Governments must fund those complements rather than treating lightweight models as replacements for them.

The World Bank supports Mission 300, which aims to provide electricity access to 300 million people in Sub-Saharan Africa by 2030. Progress on efforts like this will determine whether digital services reach rural schools, clinics, and businesses reliably.

Skills investment should extend beyond model developers. Procurement officers need to assess technical claims. Teachers and health workers need to recognize system limits. Regulators need methods for investigating harm. Local researchers need access to data and evaluation resources.

If these capabilities grow alongside deployments, small AI can become part of institutional capacity. If outside contractors retain all technical control, projects may disappear when funding ends.

The third signal is whether countries preserve choice among providers. Public agencies should demand portable data, documented interfaces, audit access, and practical exit plans. They should avoid systems that cannot operate without one vendor’s cloud or proprietary workflow.

Open technologies can help, but openness alone is not enough. Governments need organizations capable of maintaining them. They also need security practices, clear accountability, and long-term funding.

Vendor diversity would strengthen the strategy by keeping local adaptation competitive. Growing dependence on a few global providers would weaken it, even if the deployed models remain small.

These signals will become visible through procurement decisions, independent evaluations, national AI plans, and public-service results. Model announcements will be less informative than evidence that a clinic, school, or farm service still works after its pilot period ends.

The World Bank has identified a credible alternative to treating frontier-model development as the only path into AI. Its strongest insight is that usefulness depends on context, delivery, and institutional readiness.

The harder question is whether governments and funders will support the unglamorous work behind that insight. Local datasets must be maintained. Frontline workers need training. Systems require audits, updates, and appeal processes. Infrastructure still needs investment.

Readers should watch for deployments that publish measurable outcomes and give local institutions lasting control. Those projects will reveal whether World Bank small AI becomes a practical development strategy or another cycle of promising pilots. The next meaningful milestone is not a larger model. It is a smaller system that keeps working, reaches excluded users, and produces results that independent evaluators can verify.