World Bank makes ‘small AI’ a development strategy, not a smaller-model race

The story

The World Bank is closing its 2026 Annual Meetings with a deliberately different answer to the global AI race: most developing economies should not begin by trying to reproduce the largest frontier models. They should start with narrowly focused systems that work in local languages, on ordinary phones and under limited connectivity, then build capability in stages.

Paschal Donohoe, the Bank’s chief knowledge officer, sharpened that message in an interview published October 18, arguing that locally adapted “small AI” could compress long periods of economic development. The term does not describe a single technical architecture or parameter threshold. In the Bank’s usage, it covers specialised tools designed around a particular task, population and infrastructure constraint rather than an all-purpose model served from a giant data centre.

The strategy comes from the World Development Report 2026, released in August and repeatedly placed at the centre of this week’s meetings in Bangkok. The report proposes an ‘adopt, adapt and advance’ sequence: use tools that already exist, reshape them for local conditions and move toward frontier development only as skills, data, infrastructure and institutions mature. The sequencing is intended to prevent scarce capital being consumed by prestige computing projects before countries can reliably deliver basic digital services.

Its labour-market numbers challenge the assumption that generative AI will affect every economy in the same way. The Bank estimates that 4.5% of existing jobs in low- and middle-income countries face automation risk, compared with 14.2% in high-income countries. It says 16.2% of jobs in developing economies could receive a meaningful productivity lift, close to the 18.7% estimate for wealthier countries. Those are exposure estimates, not forecasts of jobs created or lost, but they point toward augmentation as the nearer-term opportunity.

Agriculture shows what the approach is meant to look like. At an official Annual Meetings session, Google described AgriConnect tools that let farmers ask spoken questions about crops, credit and markets. Google.org announced $2.5 million in catalytic funding to help the World Bank Group expand the programme to one million additional farmers by the end of 2027. The planned countries are India, Nigeria, Kenya, Ethiopia, Brazil and the Philippines. The model is built on digital public infrastructure rather than a standalone chatbot dropped into a community.

The constraint is that AI cannot compensate for missing foundations. World Bank President Ajay Banga said on October 16 that practical systems still require electricity, connectivity, device access and safeguards. The Bank’s report notes that nearly one-third of rural schools in sub-Saharan Africa lack reliable electricity and more than two-thirds lack dependable internet. A phone-level model can reduce computing and bandwidth requirements, but it cannot charge a device, create trusted records or supply a functioning institution around the advice it gives.

Localization is equally demanding. Translating an interface is not the same as validating agricultural guidance for a specific crop, medical advice against a national treatment protocol or an education system against its curriculum. Small datasets can be cheaper to maintain, but they may also encode local gaps and power imbalances. Public deployments need accuracy tests across languages and regions, clear rules for sensitive data, human escalation paths and a way for users to challenge harmful outputs.

There is also a dependency question. Countries that only import closed models may become locked into a foreign provider’s pricing, policy changes and infrastructure. Building every layer domestically is rarely efficient, but governments can require interoperable data formats, portable records, transparent procurement and the ability to replace a model without rebuilding the whole public service. That is the practical meaning of adapting rather than merely adopting.

INNOVOX analysis: the World Bank’s strongest contribution is to move the definition of AI progress away from model size. A crop-advice service that works on a basic handset, in the right language and at the moment of decision may create more public value than a frontier model a country cannot afford to operate. The test, however, is not whether the system produces fluent answers. It is whether farmers improve yields or income, clinics make safer decisions, students learn more and agencies deliver services more reliably.

What to watch next is measurable deployment. AgriConnect’s expansion provides a concrete case: the one-million-farmer target should be accompanied by evidence on active use, regional accuracy, adoption by women and smallholders, changes in decisions and real economic outcomes. The broader strategy will become credible when low-connectivity performance, error reporting, privacy protection and provider independence are treated as core infrastructure — not optional features added after a pilot succeeds.

INNOVOX analysis

The useful idea is not that smaller models are automatically better. It is that the unit of innovation should be a local service that survives poor connectivity, speaks the user’s language and fits an institution’s workflow. That reframes AI competition around delivery and evidence rather than benchmark scale — but it also puts responsibility on governments and funders to build the electricity, data protection and evaluation systems those services require.

What to watch

Watch whether AgriConnect and similar programmes publish outcome data rather than user counts; whether tools work offline or on basic phones; how local-language accuracy, privacy and appeals are tested; and whether countries can switch providers without losing data, models or public-service continuity.