Building AI Systems Where Infrastructure Fails: Lessons from 20 Years in the UN
Lessons from building AI systems across UNHCR, UNICEF, UNOCHA, and ESCWA in environments with unreliable infrastructure, fragmented data, and multilingual users.
Published 2026-05-09 · By Shahzad Asghar
Most AI systems assume reliable internet, clean data, and English-speaking users. In the United Nations, none of those assumptions hold.
I have spent 20 years building technology systems in environments where the power cuts out, the network drops to 2G, the users speak languages that most models have never trained on, and the data arrives in handwritten forms photographed on a phone. This is not an edge case. For the majority of the world's population served by UN agencies, this is the default condition.
This article shares what I have learned about building AI systems that work when infrastructure fails.
The IVR System: 700,000 Refugees Served
In 2018, UNHCR Jordan needed a way for refugees to book appointments for registration, resettlement interviews, and assistance inquiries. The challenge: most refugees did not have smartphones, reliable internet, or the ability to visit an office in person.
I led the design and deployment of an Interactive Voice Response (IVR) system that allowed refugees to call a short code, navigate a voice menu in Arabic, and book appointments without any internet connection, without any app, and without any in-person visit.
The system served over 700,000 refugees. It was replicated in five countries: Egypt, Iraq, Syria, Iran, and Ethiopia. Registration service time dropped from 3 hours per family to 45 minutes, an 83% reduction.
What made this work was not the technology. IVR is not new. What made it work was designing the system around the constraints rather than despite them. No internet required. No smartphone required. No literacy required. The refugee speaks, the system responds.
Azraq Camp: Offline-First AI
Azraq refugee camp in Jordan houses over 35,000 people. Internet connectivity is intermittent. Power is rationed. Staff devices are shared.
When I expanded the IVR system to Azraq and Zaatari camps, I had to rethink assumptions that urban deployments take for granted. The system needed to work offline, sync when connectivity returned, handle duplicate records across fragmented databases, and operate on devices with limited storage.
We reached over 90,000 additional refugees through this expansion. The key design decision was treating connectivity as a luxury, not a requirement. Every component had a graceful degradation path.
Data Integrity: 63,000 Missing Contacts
Before deploying any AI system, I led a forensic data integrity analysis of the refugee registration database. The results were sobering: 63,000 missing contact records, duplicate entries across multiple systems, and records that had not been verified in years.
This analysis resulted in 165,000 records being inactivated and 27,000 individuals being offboarded from assistance programmes. It was unglamorous work. It involved no machine learning, no neural networks, no large language models. But it was the most important AI governance work I have done, because every AI system built on that database would have inherited those errors.
The lesson: data governance precedes AI governance. Always.
Education System Integration: 90% Reduction in Delays
One of the most impactful projects I led connected UNHCR's refugee registration system with Jordan's Ministry of Education and UNESCO's Education Management Information System (EMIS). The challenge was matching refugee children across three incompatible databases with different naming conventions, different ID formats, and different data structures.
The integration reduced school enrollment delays by 90%. Refugee children who previously waited weeks or months for enrollment verification could be processed in days.
This project taught me that the hardest AI problems in the UN are not algorithmic. They are interoperability problems. Getting systems to talk to each other across organizational boundaries, data standards, and political sensitivities is where the real work happens.
Digital AAP: Selected for UN Global Pulse Accelerator
Digital AAP is an AI-enabled accountability platform I designed that processes community feedback from displaced populations. It uses natural language processing to transcribe, classify, and route voice messages from refugees to the responsible operational teams.
The platform was selected for the UN Global Pulse Accelerator programme, which validates innovations with potential for system-wide impact.
What made Digital AAP different from other NLP projects was its design philosophy: the system was built for environments where feedback arrives in multiple languages and dialects, where internet connectivity is unreliable, and where the people providing feedback cannot fill out forms.
Cybersecurity Governance
At UNHCR Jordan, I established the operation's first cybersecurity governance framework. This covered data classification, access controls, incident response, penetration testing, and vendor management.
I also designed and delivered regional cybersecurity workshops across the MENA region. Cybersecurity in a humanitarian operation is different from cybersecurity in a corporation. The threat model includes state actors interested in refugee data, the attack surface includes shared devices in field offices, and the compliance requirements span multiple jurisdictions.
What I Call Last-Mile AI
I use the term Last-Mile AI to describe a delivery philosophy: AI systems matter most when they operate reliably in the environments with the greatest constraints. Fragmented data, multiple languages, limited infrastructure, high accountability requirements.
Last-Mile AI is not a framework or a product. It is a set of design principles: build for the constraint, not the ideal. Test in the field, not the lab. Govern the data before governing the model. Measure outcomes for the people served, not for the organization deploying.
For more about my work, visit my professional profile or explore my full project portfolio.