World Development Report 2026 (World Bank, August 2026): under a tenth of developing-economy jobs face AI automation

Low- and middle-income countries have 4.5% of existing jobs at risk, compared with 14.2% in high-income countries. The report also estimates that AI could meaningfully improve productivity in 16.2% of jobs in developing economies. Its sequencing advice is to use existing tools first, adapt them to local languages and data, and build frontier systems only where resources allow. Countries should first use existing AI tools, then modify them for local languages, institutions, data and requirements. Building frontier AI capabilities should come later and is relevant mainly to countries with sufficient resources. Adaptation matters because models built elsewhere often underperform locally. A model developed in the UK to assess suspected COVID-19 patients performed worse when tested on Vietnamese patient data. A diabetic-eye screening tool rejected more than 20% of images taken by nurses because clinic lighting differed. Meanwhile, many companies now offer large language models that are freely available or open for modification, including Apertus, DeepSeek, Falcon, Gemma, Kimi, Mistral, Llama, and Qwen. Public deployments are already visible, such as AI-supported weather forecasts for more than 30 million farmers in India. For a company building AI products, the report points to localization work, including language coverage, local data and institutional fit, as where value sits rather than in model scale. The same report notes that advanced chips, data centres and frontier models are concentrated in a few firms and countries, creating dependency risks. Teams selling into these markets should plan for buyers who weigh that concentration when choosing vendors.