What Is Agentic AI and Why It Matters for Humanitarian Operations
An explanation of agentic AI and its practical application in UN humanitarian operations, with real project examples from UNHCR and ESCWA.
Published 2026-05-09 · By Shahzad Asghar
Agentic AI refers to artificial intelligence systems that can plan, reason, use tools, and take autonomous action toward a defined goal. Unlike traditional AI models that respond to a single prompt and stop, agentic systems execute multi-step workflows, make decisions at each stage, and adapt based on outcomes.
In humanitarian operations, this distinction matters. The difference between a chatbot that answers questions and an agentic system that receives a voice message, transcribes it, classifies its urgency, routes it to the right team, and logs it in a case management system is the difference between a tool and an operational capability.
I have spent the past several years building agentic AI systems for the United Nations. This article explains what agentic AI is, why it matters for humanitarian work, and what I have learned from deploying these systems in regulated environments.
How Agentic AI Differs from Traditional AI
Traditional AI systems are reactive. You give them an input, they produce an output. A language model answers a question. An image classifier labels a photo. A recommendation engine suggests a document.
Agentic AI systems are proactive. They receive a goal, break it into steps, decide which tools to use at each step, execute those steps, evaluate the results, and adjust their approach if needed. They operate as autonomous agents within defined boundaries.
The key architectural patterns include:
Multi-agent orchestration where multiple specialized agents collaborate on a task. One agent handles transcription, another handles classification, a third handles routing. An orchestrator coordinates their work.
Tool-using agents that can call external APIs, query databases, send notifications, and interact with other systems. The agent decides which tools to use based on the task requirements.
Retrieval-augmented generation (RAG) where agents query document stores, knowledge bases, or operational databases before generating responses. This grounds their output in verified information rather than model assumptions.
Human-in-the-loop safeguards where agents escalate decisions to human reviewers when confidence is low, when the decision involves sensitive data, or when policy requires human approval.
Agentic AI Projects I Have Led
Digital AAP: Multi-Agent Feedback Processing
Digital AAP is an AI-enabled accountability platform I designed for UNHCR that was selected for the UN Global Pulse Accelerator. The system uses multiple agents to process community feedback from displaced populations.
One agent transcribes voice messages. Another classifies them by topic and urgency. A third routes them to the responsible operational team. A fourth generates management reports from aggregated feedback patterns. Each agent operates autonomously within its domain, and an orchestration layer coordinates their work.
Voice-First Refugee Feedback System
I developed an AI-enabled voice feedback solution for UNHCR that captures spoken messages from refugees, transcribes them across multiple languages, classifies them, and routes them for follow-up. The system served over 700,000 refugees across five countries.
This is agentic AI at its most practical. The refugee speaks. The system listens, understands, decides, and acts. No forms. No apps. No internet required on the user's end.
AI Governance Validation Agent
I designed an agentic workflow that audits other AI systems for policy compliance. The agent reviews system configurations, data handling practices, and decision logs against governance requirements. When it finds a gap, it generates a specific remediation recommendation.
This is AI auditing AI, with human oversight at the review stage.
Why This Matters for the UN System
The United Nations operates in environments where staff are stretched thin, data arrives in fragments, and operational decisions must be made quickly under uncertainty. Agentic AI does not replace human judgment. It handles the mechanical parts of complex workflows so that human expertise can focus on the decisions that require it.
For a complete overview of my agentic AI work, visit my agentic AI resource page. To see the full project portfolio, visit AI projects.
What I Have Learned
Three lessons from deploying agentic AI in humanitarian settings:
Governance must be built into the agent, not applied after deployment. Every agent I build has explicit boundaries, escalation rules, and audit logging. An autonomous agent without governance is a liability.
Start with the workflow, not the technology. The best agentic AI systems automate workflows that humans already perform well but cannot scale. Understand the workflow first, then design the agents.
Human-in-the-loop is not optional in humanitarian contexts. Every agentic system I deploy includes mandatory human review points for decisions that affect vulnerable populations. Autonomy is a spectrum, not a switch.