AI Projects in Humanitarian and Development Operations

A portfolio of artificial intelligence projects built and deployed across United Nations humanitarian and development operations. These are working systems, not pilots on paper: voice-based refugee feedback, an interactive voice response appointment platform serving more than 700,000 refugees, data integrity analysis that restored trust in core registration systems, and agentic AI for decision support. Each was delivered under real infrastructure, language, governance, and operational constraints, the conditions that decide whether an AI system is trusted in the field.

DigitalAAP — Voice-Based Refugee Feedback

Problem. Affected populations often have no reliable, dignified way to give feedback to the organisations that serve them, and the feedback that does exist is hard to analyse at scale or across many languages.

Approach. DigitalAAP is an AI feedback platform that collects community input through accessible channels and uses natural language processing to classify, route, and summarise it, while protecting the identity of the people who provide it. The design treats data protection and consent as the starting point rather than a final compliance check.

Outcome. DigitalAAP was selected for the United Nations Global Pulse Accelerator programme, recognising its potential to make accountability to affected populations a routine, measurable practice rather than an afterthought.

Why it matters. Accountability to affected populations is often promised and rarely measured, because the feedback arrives faster than anyone can read it and in more languages than any team can cover. By turning unstructured feedback into structured, classified signal, the system lets an operation see patterns early, respond to them, and show that it listened. The hard part was never the model. It was building something dignified, multilingual, and safe enough that people would trust it with their concerns.

IVR Appointment System — 700,000+ Refugees

Problem. Appointment booking by physical queue is slow, costly, and exposed to abuse. It excludes people with limited literacy, limited connectivity, or no smartphone, exactly the people humanitarian services exist to reach.

Approach. An interactive voice response system lets refugees book and manage appointments by phone. It is voice-first and multilingual, designed for low-bandwidth conditions and basic devices, so access does not depend on owning the right technology.

Outcome. The system served more than 700,000 refugees, achieved an 83% reduction in registration service time, and was replicated across five country operations including Egypt, Iraq, Syria, Iran, and Ethiopia. It became UNHCR's first global IVR appointment platform.

Why it matters. The replication across five countries is the part that proves the design. A tool that works once can be luck; a tool that transfers across different operations, languages, and infrastructure has solved the real problem rather than a local one. By removing the physical queue, the system also removed the travel, the lost day of work, and the exposure that came with it, which matters most for the people least able to absorb those costs. It is a direct example of delivering AI and automation at the last mile.

Data Analytics and Integrity

Problem. Large refugee registration databases accumulate duplicates, gaps, and errors that erode confidence in the data and in every service built on top of it.

Approach. Applied data analytics and forensic integrity analysis to detect duplicates, recover missing links between records, and integrate external systems through secure APIs, so that operational decisions rest on data people can trust.

Outcome. The work identified 63,000 missing contacts through data integrity analysis and reduced school enrolment delays by 90% by integrating Ministry of Education systems into the enrolment workflow. It was delivered as part of a wider 2.5 million dollar digital transformation programme.

Why it matters. Every AI system on this page depends on the data underneath it, which is why integrity work comes first. A model trained or run on a database full of duplicates and broken links will produce confident, wrong answers, and in a protection context a wrong answer has consequences. Recovering tens of thousands of missing contacts is not a back-office task; it is what makes the difference between a family being reached and being lost in the system. Clean, connected data is the unglamorous foundation that the visible AI sits on.

Agentic AI and Decision Support

Problem. Many operational decisions depend on knowledge scattered across documents and systems that staff cannot search quickly under pressure.

Approach. Agentic AI systems using retrieval-augmented generation and multi-agent architectures bring the right information to the point of decision. Human oversight and clear escalation are designed in, so the system informs a decision rather than replacing the person accountable for it.

Outcome. Decision-support tools and project assistants that turn institutional knowledge into something usable at the moment it is needed, with the audit trail and human control that regulated environments require. One example is the ADP portal at UNESCWA, which applies AI to the automatic analysis of charts and data so analysts spend less time formatting and more time deciding.

Why it matters. Agentic AI is where the most value and the most risk sit together. A system that can act, not just answer, has to be held to a higher standard of oversight, logging, and escalation, because the cost of an unchecked action is real. The design principle across these tools is consistent: let the AI carry the search, the retrieval, and the drafting, but keep a person accountable for the decision. That balance is what makes autonomy acceptable in a regulated institution rather than a liability.

One Method Behind the Portfolio

These projects share a single delivery method, Last-Mile AI: building systems that work under the real constraints of the places they serve, where connectivity is thin, languages are many, data is sensitive, and the consequences of failure are carried by people who can least afford them. They are governed with the discipline of the NIST AI Risk Management Framework and the wider practice of AI governance in the United Nations, so capability and control advance together rather than one racing ahead of the other. The common thread is not a single algorithm but a way of working: start from the operating reality, protect the people in the data, keep a human in command, and measure the result in outcomes rather than output. For the wider economic context, see the economic impact of AI on development.