Compact AI models that run on a mobile phone with a weak signal could accelerate growth in emerging economies within decades rather than centuries. The position was outlined by Paschal Donohoe, chief knowledge officer of the World Bank Group, in comments reported by the Guardian. The proposal matters because it reframes AI adoption for poorer nations around local languages, low connectivity and single-task tools.
Small models for weak connections
Donohoe described "small AI" as specialised models trained on narrower data for one task. Such tools are designed to work on phones, with weak connections and in local languages. He linked the approach to the bank's AgriConnect scheme in India, where farmers need soil analysis in their own language. In his assessment, there will not be enough scientists and agricultural experts to serve every small farm individually.
The delivery model rests on cooperation between the World Bank, governments and private companies. Products would be built to fit local needs rather than transferred unchanged from richer markets. The emphasis is on task-specific assistants that function without stable broadband or high-end devices. That design choice lowers infrastructure demands while keeping the use case narrow and measurable.
AI was a central theme at the World Bank and IMF meetings held this month in Bangkok. IMF chief Kristalina Georgieva said governments must first provide power, digital infrastructure and protection from bad actors. She argued that with those basics in place, adoption can become as democratic as the cellphone. The World Bank's August development report underlines the gap: almost a third of rural schools in sub-Saharan Africa had no electricity.
What small AI means for business
For companies operating in emerging markets, the signal is a shift toward lightweight, offline-tolerant applications in agriculture, advisory services and field operations. Vendors that package one clear function, local-language support and low data use will fit procurement by development agencies and governments. Large organisations can pilot such tools inside existing extension or distribution networks, while smaller firms can build niche products around one crop, region or workflow.
The constraints remain substantial and should shape any buying decision. Power supply, connectivity and device capability still determine whether a model works in practice, not its benchmark score. Georgieva warned that 40% of jobs worldwide are likely to be changed or wiped out by AI over time, with entry-level roles most exposed. Buyers should therefore verify language quality, offline behaviour, data handling and who maintains the model after deployment.
The marker to watch is financing and follow-through after Bangkok. Donohoe said the World Bank had raised $120bn in private capital over the past year as the US and UK cut aid budgets. If that capital converts into funded deployments such as AgriConnect, small AI will move from agenda item to procurement pipeline.
