AI for Humanitarian Organisations
Last updated:
Humanitarian AI is the use of artificial intelligence in aid operations, covering registration, feedback, targeting, forecasting, and service delivery for people affected by crisis. Humanitarian AI succeeds or fails on conditions that rarely appear in a demonstration: intermittent connectivity, many low-resource languages, sensitive personal data, and people who cannot appeal a wrong decision.
A practitioner guide to adopting artificial intelligence in humanitarian organisations, written from delivery rather than from policy. It covers the use cases that survive contact with the field, the risks that matter most when the people in the data are displaced, and the governance that keeps a human accountable for consequential decisions.
What humanitarian AI actually means
Humanitarian AI covers the systems an aid organisation runs on behalf of people in crisis: registering them, scheduling their appointments, hearing their complaints, forecasting where they will move, and deciding who receives assistance first. The subject is not the model. The subject is a person who cannot choose another provider, cannot easily appeal, and carries the cost when a system is wrong.
That asymmetry is what separates humanitarian AI from commercial AI. A recommendation engine that misfires loses a sale. A targeting model that misfires removes a family from an assistance list. The engineering is similar; the tolerance for error, and the obligation to explain it, is not.
Where AI holds in humanitarian operations
Four categories have repeatedly worked in practice. Access and scheduling: voice and messaging systems that let people reach a service without an app or a data plan. Feedback and accountability: transcribing and classifying what affected people say, so complaints reach the team that can act rather than dying in a spreadsheet. Data integrity: finding duplicates, gaps, and errors in registration data that quietly degrade every service built on top of it. Staff productivity: retrieval over project documents and policy so officers stop rediscovering the same answer.
What these share is that the AI sits behind an operational process with a named owner, not in front of a beneficiary making an unreviewable decision. See the AI for refugee operations hub for the delivered examples, including an interactive voice response appointment system used by more than 700,000 refugees.
Where it fails
Most humanitarian AI failures are not model failures. They are delivery failures: the system assumed connectivity that was not there, a language it did not support, a governance step that was skipped, or a user context it never accounted for. That pattern is set out in full in the Last-Mile AI Framework, which treats those constraints as design inputs rather than deployment surprises.
Three failure modes recur. A pilot proves a model and never confronts the operating environment. A chatbot is placed in front of crisis-affected people without a route to a human. And a system is procured rather than built, so nobody inside the organisation can explain how it decides or switch it off.
The governance that has to come with it
Governance in this sector is not a statement of principles. It is a set of answers to operational questions: who may approve a model for production, who can switch it off, how risk is tiered by consequence rather than by technology, and what observable condition brings a system down. Those answers are set out in the AI governance playbook, with a model register and approval gate you can use unchanged.
Data responsibility carries the same weight. Humanitarian data describes people who may be at risk from the state they fled, so data minimisation, retention limits, and access control are protection measures rather than compliance paperwork. For the wider institutional view, see AI governance in the United Nations.
How to decide whether a use case is suitable
Four questions settle most cases. What happens to a person if the system is wrong, and can they appeal? Does the organisation hold data of sufficient quality to support the task at all? Is there a named owner who will carry the decision, not a committee? And is there a channel to a human at the point where the system stops being confident? A use case that cannot answer all four is a research project, not a deployment. The use case discovery playbook turns those questions into a scoring sheet.
For related work, see agentic AI in humanitarian operations, the delivered project portfolio, and AI security and assurance.
Related reading
These go deeper on parts of this subject:
AI chatbots for crisis-affected people — where a conversational interface helps and where it must not be the only route.
AI and gender-based violence case management — the highest-sensitivity data an aid organisation holds.
Classifying refugee call transcripts — a delivered system, with the method written out.
A practical AI roadmap for national NGOs — sequencing for organisations without a data team.
Common questions
How can AI be used in humanitarian operations?
Four categories have repeatedly worked in practice. Access and scheduling systems, such as voice or messaging services that let people reach a service without an app or a data plan. Feedback and accountability systems that transcribe and classify what affected people say so complaints reach the team that can act. Data integrity work that finds duplicates, gaps, and errors in registration data. And staff productivity tools such as retrieval over project documents and policy. What these share is that the AI sits behind an operational process with a named owner rather than in front of a beneficiary making an unreviewable decision.
What are the risks of AI in humanitarian organizations?
The defining risk is asymmetry: the people in the data cannot choose another provider, cannot easily appeal a decision, and carry the cost when a system is wrong. A targeting model that misfires can remove a family from an assistance list. Humanitarian data also describes people who may be at risk from the state they fled, so data exposure is a protection risk rather than a privacy inconvenience. Practical risks include deploying a chatbot without a route to a human, and procuring a system nobody inside the organisation can explain or switch off.
What is responsible AI in the humanitarian sector?
Responsible AI in this sector means control keeping pace with capability. In operational terms: documented use cases, a named owner who can approve and withdraw a system, risk tiered by what happens to an affected person when the system is wrong rather than by which technology is used, data minimisation and retention limits treated as protection measures, human oversight at consequential decisions, and monitoring that continues after launch with a written stopping condition.
How should NGOs assess whether an AI use case is suitable?
Four questions settle most cases. What happens to a person if the system is wrong, and can they appeal? Does the organisation hold data of sufficient quality to support the task at all? Is there a named owner who will carry the decision rather than a committee? And is there a channel to a human at the point where the system stops being confident? A use case that cannot answer all four is a research project rather than a deployment.
Why do humanitarian AI pilots fail?
Most failures are delivery failures rather than model failures. The system assumed connectivity that was not there, a language it did not support, a governance step that was skipped, or a user context it never accounted for. A common pattern is a pilot that proves a model under good conditions and never confronts the operating environment where it must actually run.
How do humanitarian agencies protect data when using AI?
By treating data minimisation, retention limits, and access control as protection measures rather than compliance paperwork, because humanitarian data describes people who may be at risk from the state they fled. In practice this means collecting only what the task requires, deciding where inference happens and who can see the prompts, deleting on a schedule, and keeping sensitive workloads under direct institutional control rather than sending them to a third party by default.
Get the next essay by email
One practical essay a month on AI governance, agentic AI, and digital delivery in the UN system. No marketing, no forwarding of your address.