Shahzad Asghar · Head of Data and Digital Solutions, UN-ESCWA
AI governance and digital transformation in the United Nations
Last updated:
I am Shahzad Asghar, a senior United Nations AI and digital transformation leader with over 20 years across UNESCWA, UNHCR, UNICEF, UNOCHA, CIDA, and public-sector institutions. I help UN agencies and member states deploy AI that is accountable, secure, and useful in the field. My work sits at the intersection of AI governance, large language model applications, cybersecurity, enterprise data systems, and operational modernization in multilingual, high-risk, and heavily regulated environments.
Key takeaways
- Most AI failures in humanitarian and public-sector operations are delivery failures, not model failures. The system assumed connectivity, a language, or a governance step that was not there.
- The Last-Mile AI Framework treats those constraints as design inputs. It is published as a citable preprint (DOI 10.5281/zenodo.21921490).
- Governance works when it produces decisions rather than principles. The six AI playbooks give the rules, templates, and failure signals.
- Evidence from the field: an interactive voice system reached more than 700,000 refugees, and a data-integrity exercise recovered 63,000 missing contact records.
AI leadership in real-world operating environments
I have led digital and AI initiatives across humanitarian, development, and institutional settings where reliability, privacy, governance, and operational value all matter at once. That work spans enterprise data platforms, AI-enabled service delivery, secure cloud migration, cybersecurity governance, and digital systems designed for multilingual and high-risk contexts.
Not everyone approaches this the same way. One school of thought argues that institutions should wait for regulation to settle before deploying AI at scale, and there is a reasonable case for caution when the people affected cannot appeal a decision. Another argues that waiting concedes the ground entirely, since AI arrives inside purchased software whether or not you have a policy for it. My own position sits between the two: adopt deliberately, govern from the first day, and keep a human accountable at the points where a wrong output touches a person. According to the NIST AI Risk Management Framework, risk should be assessed by context and consequence rather than by technology, which is the same conclusion field experience produces.
Key terms, defined
- Last-Mile AI
- The practice of delivering AI that works in the real conditions where it is used, not the controlled conditions where it is built. Read the full definition.
- Agentic AI
- Systems that plan across steps, call tools, and act toward a goal rather than returning one output per input. See humanitarian applications.
- AI governance
- The decision rights, risk tiering, and stopping conditions that determine who may deploy a model and who switches it off. How it works in the UN system.
- Sovereign AI
- A country's ability to develop, run, and govern AI on its own terms: control over data, compute, models, talent, and rules. Strategy and cost.
Selected impact, with sources
| Result | System | Where and when |
|---|---|---|
| 700,000+ refugees served | UNHCR's first global interactive voice response appointment system | UNHCR Jordan, 2018–2024 |
| 83% reduction in service time | Refugee registration process redesign | UNHCR Jordan |
| 63,000 missing contacts recovered | Data integrity and forensic analysis | UNHCR Jordan |
| 90% reduction in enrolment delays | Ministry of Education systems integration | Jordan |
| Replicated across 5 countries | Egypt, Iraq, Syria, Iran, Ethiopia | UNHCR operations |
| Selected for the UN Global Pulse Accelerator | DigitalAAP, an AI-driven refugee feedback platform | Third cohort |
These figures come from programmes I led directly. The methodology behind them is documented in the citable preprint Last-Mile AI: A Practitioner Framework (Zenodo, 2026) and verified against my ORCID record.
How to deploy AI where infrastructure fails
If you are starting an AI programme in a constrained environment, this is the sequence I would follow. Each step exists because skipping it is what I have seen fail.
- Start from the operating reality. Ask what has to be true for the system to be trusted where it will actually run, before choosing any model or vendor.
- Decide the resolution question. Most operational decisions need patterns, not persons. Producing person-level data you do not need is liability, not thoroughness.
- Name the accountable owner. Write down who accepts the output and who is authorised to switch the system off, with the observable condition that triggers it.
- Test in your working languages. A model that performs well in English can fail in Arabic dialects, Urdu, or local languages. Test dialect by dialect, because your users do not speak the average.
- Put a human at the points of consequence. Not everywhere, which kills the value, but wherever a wrong output touches a person.
- Monitor after launch. Assign a review date at approval. Systems drift as data, vendors, and context change.
You can work through each of these in depth in the AI playbooks, which include the model register schema, the consequence tiering criteria, and the approval gate checklist as templates you can use without modification.
Recent Writing
Frequently Asked Questions
Who is Shahzad Asghar?
Shahzad Asghar is a senior United Nations AI, data, cybersecurity, and digital transformation leader with over 20 years of experience across multiple UN agencies and public-sector institutions.
What AI work has Shahzad Asghar led?
He has led applied AI work in accountability systems, voice-based feedback, policy and decision support, workflow automation, knowledge platforms, and regulated operational environments.
What is Last-Mile AI?
Last-Mile AI is Shahzad's delivery approach for building practical AI systems that work in real-world environments with infrastructure, language, governance, and operational constraints.
What sectors has Shahzad worked in?
His work spans the United Nations, humanitarian operations, development settings, government partnerships, and regulated public-sector institutions.
What areas does Shahzad specialize in?
He specializes in AI governance, LLM applications, cybersecurity, digital transformation, enterprise data systems, and operational decision support.
How does the United Nations govern AI?
UN agencies govern AI through institutional risk frameworks, human-in-the-loop controls, data protection rules, and accountability structures. Shahzad explains the practice in AI Governance in the United Nations, based on direct implementation experience.
Where is the Last-Mile AI Framework published?
The framework is published as a citable preprint: Asghar, S. (2026), Last-Mile AI: A Practitioner Framework for Delivering Artificial Intelligence Under Infrastructure, Language, and Governance Constraints, Zenodo, doi.org/10.5281/zenodo.21921490. The full framework is at shahzadasghar.com/framework.
Publications
- Asghar, S. (2026). Last-Mile AI: A Practitioner Framework for Delivering Artificial Intelligence Under Infrastructure, Language, and Governance Constraints. Zenodo. doi.org/10.5281/zenodo.21921490
- Asghar, S. (2026). When the Map Knows Where You Sleep: The Two Dilemmas of Humanitarian GeoAI. Medium
Standards and reference material I work from include the NIST AI Risk Management Framework, the UNESCO Recommendation on the Ethics of AI, and the OECD AI Principles.
Where to start
If you are responsible for AI in an institution, start with the problem you already have rather than the tool you have heard about. Three practical entry points, depending on where you are:
- You need governance that produces decisions. Read the AI Governance Playbook and take the model register schema and approval gate as they are.
- You are choosing what to build first. Use the four-axis scoring model in the use case discovery playbook.
- Your team needs to learn the material. The AI Learning Hub is free and self-paced, and the Last-Mile AI Framework explains the delivery method behind all of it.
My verdict after twenty years: the model is rarely the hard part. Governance, language coverage, and the last mile decide whether a system helps anyone. If you are working on something in this space, I would like to hear about it — get in touch, or read more about my background.