1. Introduction: AI-Powered Anticipatory Action

GiveDirectly has used an AI-based flood prediction tool to deliver anticipatory cash assistance to at-risk communities in Nigeria and Bangladesh. By incorporating flood forecasting data from Google Research and funding from Google.org, this application identifies forecasted flood events at a granular level — targeting specific villages rather than whole regions. This precision allows GiveDirectly to design timely financial support through cash payments before shocks strike, enabling communities to better prepare and cope with impending floods. Through this proactive approach, community resilience and preparedness are increased, with the goal to preserve human dignity, life, and livelihoods and to reduce the cost of post-disaster assistance by mitigating the impact of floods.

2. Background: Why Anticipatory Cash?

Flood events often hit affected populations without the necessary resources to prepare and respond effectively, leading to significant economic and social hardships. Homes, crops, and infrastructure can be devastated. Cash assistance is ideally provided in advance of the flood happening, to allow communities to prepare — for example, by purchasing provisions, making changes to buildings, or evacuating. However, for anticipatory cash aid to be effective and sustainable, it must be accurately targeted to ensure it reaches the most vulnerable populations. Likewise, post-disaster cash responses must focus on those worst affected, ideally identified at a micro, community-specific level.

3. AI Development: Google Research and Local Calibration

GiveDirectly collaborated with Google Research and JBA Global Resilience to deploy this AI-based application. Through Google’s Flood Hub, the team receives essential hydrological and inundation forecasts. The hydrological forecasts provide data on river levels and water flow, while the inundation forecasts map out areas likely to experience flooding. Compared to other forecasting services such as GloFAS and GeoGlows, Google Flood Forecast was chosen for its ability to deliver detailed insights at the community level. JBA Global Resilience then developed a trigger system that provided daily flood risk updates to GiveDirectly based on the forecasts.

4. Operationalizing AI: From Trigger to Payment

The AI-supported forecasting and trigger generation is the first step in the end-to-end process of cash assistance. Country program teams receive a list of newly triggered communities every morning and follow a standard operating procedure called “trigger to payment process” by which they request the technology team to release cash payments. Because GiveDirectly pre-enrolls individuals for anticipatory action, the payment can reach them in as quick as 48 hours, as the team has demonstrated in Nigeria.

Since the process of generating triggers based on Google’s Flood Hub is automated through an API, ongoing support is a “low lift” from a technology point of view once a country setup has been completed. Based on learnings from Bangladesh and Nigeria, regularly reviewing and adapting the triggers is important to ensure ongoing accuracy. When adding new countries, triggers need to be fine-tuned to that new context, because floods behave differently in different settings.

5. Key Learnings: Local Data + AI = Greater Accuracy

6. Future Plans: Scaling to Drought and New Countries

GiveDirectly plans to extend the AI application to forecast other disaster events, such as droughts, and to expand to other countries, such as Kenya, and ideally more countries subject to funding and availability of Google’s Flood Forecasts. The team will keep supporting existing projects and assessing what benefits they are delivering. One of their aims is to build trust in this approach to anticipatory action so that governments, humanitarian organizations, and donors will consider doing the same.

7. Frequently Asked Questions (FAQ)