Last-Mile AI: A Practitioner Framework for Delivering Artificial Intelligence Under Infrastructure, Language, and Governance Constraints

Most operational failures of artificial intelligence systems in humanitarian and public-sector settings are not model failures. They are delivery failures: the system assumed connectivity that was not there, a language it did not support, a literacy level its users did not have, or a governance step that was skipped under deadline pressure. This paper presents Last-Mile AI, a practitioner framework developed across two decades of United Nations field operations, structured around five principles: context before code, design for failure, human loop always, local ownership first, and measure what matters. We describe three applications from UNHCR Jordan (2018-2024): an interactive voice response appointment system, UNHCR's first at global level, which served more than 700,000 refugees and contributed to an 83% reduction in registration service time within a USD 2.5 million digital transformation programme; the DigitalAAP feedback platform, selected for the third cohort of the UN Global Pulse Accelerator; and a data integrity analysis that recovered 63,000 missing contact records. Solutions were replicated in Egypt, Iraq, Syria, Iran, and Ethiopia. We then extend the framework to agentic AI systems and discuss its limitations as single-organization operational evidence.