Five things we've learned using mobile data and AI/ML to identify people in need

Summary

Cross-country overview of GiveDirectly's MobileAid programme across Togo, DRC, Malawi, and Bangladesh. Reports phone-based targeting AUC of 0.61-0.72 across countries. Details data privacy safeguards including pseudonymisation, data protection impact assessments, formal data-sharing agreements, and in-country data processing (Bangladesh data stored on government servers). Notes open-source CIDER library for ML methodology. Reports 6-18 month setup time, 11 telecoms engaged across six countries, and shift toward models where data processing happens inside telecom or government systems.