All posts
Summary
- GiveDirectly used AI flood forecasts to deliver early cash to people in Kogi, Nigeria — reaching the right households before floods peaked.
- Once forecasted flood thresholds were met, we sent $105 to over 4,600 individuals before the flood peak.
- This early cash helped families evacuate, protect assets, and rebuild — incomes more than doubled, food insecurity dropped by 90%, and 93% felt better prepared for future floods.
Getting cash to people before disaster strikes can save lives. That’s why we piloted an AI-powered flood response in Nigeria, predicting where floods would hit and delivering early cash to those most at risk.
By combining cutting-edge forecasts with local insights, we reached flood-prone communities faster, more affordably, and with greater precision than GloFAS1 — delivering aid before the waters rose.
First, we identified the most flood-vulnerable wards in Kogi, Nigeria
Kogi state is one of the most flood-affected regions in West Africa.
In May 2024, we used Google’s historical flood data, wealth indicators, and local insights from the Nigeria Hydrological Services Agency (NIHSA) to identify the 6 wards most vulnerable to flooding to run our pilot (see Appendix).
Each ward includes between 3-10 different communities for a total of 53 discrete communities. Each of these essentially becomes its own island during peak flooding because the roads are inaccessible.
Independent evaluations later confirmed our targeting was accurate. All 53 communities we targeted were flooded in October, 2024. Satellite imagery, survey data, and an IOM post-flood report validated that Ibaji — our focus area — had the highest number of affected people in the region.
Then we pre-enrolled thousands for payments ahead of floods
This modeling helped us select the right communities. But to get cash there quickly and affordably, we replaced door-to-door enrollment with a text message (USSD) that saved at least $80,000 in field staff.
- 📣 Before enrollment opened: Our field team led community meetings in every ward to explain who was eligible and made clear that pre-enrolling wasn’t a guarantee of payment.
- 📱Pre-enrollment period: People dialed a short code and self-reported age, location, ID, and bank info. 38,000 did this in two weeks but only 18,000 of those people self-reported living in one of the 6 eligible wards.
- 🪪 Remote verification: Our call center then reached out to those 18,000 potentially eligible people by phone — asking things like a local clan leader’s name or the nearest hospital to confirm they were truly local. Our team connected with 15,000 people and, through the process, verified around 5,000 recipients as eligible.
- 💪 In-person verification check: Finally, our team went into the field for an in-person check, which was done in close collaboration with community leaders. We checked the 5,000 people we verified remotely as well as the 3,000 people we were unable to reach by phone in order to be inclusive of those with limited phone-access. Ultimately we fully confirmed 4,600 as eligible2 and registered them. Every eligible person got a payment.
A traditional field team would have taken a year to reach the same number of people. This approach let us focus more funds on recipients and less on logistics.
Our hyper-accurate flood model told us when to trigger payments
We used AI-powered forecasts and community insights to design payment “triggers” for when flooding was about to peak. Most anticipatory aid programs set flood triggers at a broad Local Government Area (LGA) level.
We took a more granular approach that let us target areas up to 10x smaller by:
- Partnering with hydrology experts at JBA Global Resilience and consulting local leaders.
- Using Google’s AI-driven FloodHub forecasts to track flood risk in real time.
- Defining thresholds as 20% of a community area forecasted to flood.
- Monitoring forecasts daily and cross-checked them with community observations on water levels and flood history.
These improvements were based on failures from our 2022 pilot. This wasn’t our first anticipatory program in Kogi, Nigeria — and we came prepared with hard-won lessons:
- In 2022, too few people were pre-enrolled, limiting our flexibility. In 2024, We pre-enrolled 38,000 people and reached 3,250 with early cash.
- In 2022, we were only set up to send payments after major flooding had already hit. In 2024, we delivered aid before severe floods arrived.
While AI helped us act quickly, community input made it work. Local residents helped refine flood triggers by sharing past flood patterns and monitoring daily water levels. These insights allowed us to catch early flood events that AI alone might have missed. Without this collaboration, many individuals might have missed the narrow window for timely support.
As flood waters started rising, our payment triggers were met and we sent cash.
We sent early payments to 3,250 people between August 29 and October 6, several weeks before floods peaked on October 21 (according to NIHSA’s data and satellite imagery).
In two wards, we temporarily paused payments to resolve identity issues and ensure the right people were enrolled. Once those checks were complete, another 1,400 people received cash on October 24 — bringing the total number of individuals paid to over 4,600.
Families used this early cash to reduce losses ahead of floods and recover after
These floods were devastating for the wards we supported: 63% of recipients saw a drop in farming outputs and 51% had damaged homes. Because we got cash there early, people had the resources to protect their families, safeguard what they owned, and start rebuilding quickly.
They prepared before the flooding peaked.
Recipients used the first $105 transfer ahead of peak flooding to strengthen their defenses and evacuate safely. With early cash in hand, families were able to:
- 🛶 Build bamboo rafts to protect food, belongings, and crops
- ⛰️ Evacuate to higher ground, especially moving children and elders
- 🧺 Stock up on essentials, like food and medicine
“The transfer came at a perfect time. It helped me evacuate my properties that could have been damaged by the flood.” — 31-year-old woman
“We built rafters to store farm produce and foods. We also moved the minors to neighbors with storey buildings.” — 34-year-old man
“We evacuated the minors to upper land where it is not easily flooded, we also construct rafters to stay, cook and keep watch of other properties from thieves.” — 36 year old man
They also rebuilt after the waters receded.
A second $210 transfer helped families recover lost income, rebuild homes, and strengthen their long-term resilience:
- 💰 Household savings rose from 7% to 56%
- 📈 Weekly income more than doubled, from ₦4,500 to ₦10,000
- 🍲 Food security improved, with 97% facing food shortages before and just 7% after
- ✅ 93% of recipients felt better prepared for future floods
Remote targeting increases speed and cuts cost, but surfaced familiar barriers
This pilot leaned heavily on remote systems like USSD self-enrollment and phone-based verification for speed and simplicity. These come with tradeoffs:
🤳🏾 The most vulnerable people are the least likely to have a phone
We can only enroll people who already use a mobile phone, and phone ownership is lowest among the poorest individuals. While independent data is scarce, according to one 2016 survey of rural individuals in Kogi State, approximately 84.8% of individuals owned or had access to a GSM (mobile) phone. As we’ve mentioned elsewhere, those without access are likely to be the most vulnerable (poorest, elderly, etc), so our model risks excluding some of the people we most want to reach.
🪪 Verifying someone remotely doesn’t mean verifying them for free
USSD self-enrollment made registering fast, but confirming who was actually eligible took real people and real time — thousands of phone calls, then an in-person visit for everyone we could reach and everyone we couldn’t, before any final decision (see the verification steps above). Remote targeting shifts where the human effort goes; it doesn’t remove it.
💬 Being excluded without an explanation costs trust
Careful verification still means tens of thousands of people who pre-enrolled are not paid. How we communicate that — before, during, and after a program — matters as much as the verification itself. This is an area we’re actively working to do better: making sure people understand why they weren’t selected and not left wondering if they will get cash later on.
🌍 Geographic triggers pay by area, not by individual need
Unlike our poverty alleviation programs, our 20% flood-threshold trigger is designed to move fast across an entire verified flood zone — it doesn’t run a means test on every individual inside it. That’s a deliberate tradeoff: speed and coverage over precision targeting. It means some better-off individuals within a flooded ward received cash alongside the poorest and that poor people who may have moved to higher grounds before the program began may have incurred flood-related losses but did not benefit. Still, when we conducted randomly selected recipient surveys, we found that the median weekly income across recipients was 4500 Naira ($3 USD).
☎️ Staying reachable has its own costs
Some recipients were hard to reach during the verification window. Field teams found that people weren’t near their phones when we called (having left their phone to charge, for example), and in some poor-signal areas people had to climb trees or get onto rooftops to make a call. This on top of the fact that digital literacy is worse among the poorest community members. These gaps are why in-person follow-up mattered: it caught people the phone-based process alone couldn’t.
🛡️ Combatting impersonation is harder with remote-first programs
Impersonation is a real risk with any cash program, but especially so in a program with limited field staff. In Kogi, we referred third-party scammers charging fake “registration fees” to local authorities, gave field staff badges so recipients could tell who was really from GiveDirectly, and sent a closing SMS telling recipients the program had ended and that anyone still asking for money was a scammer. Our teams also ran ward-by-ward re-sensitization visits and exit focus groups in every community, plus close-out meetings with state, LGA, and community leadership specifically on impersonation risk. Still, this wasn’t a foolproof process, and safe community exit is something we’re still working to do better.
None of this is a reason to abandon tech-enabled targeting, but rather to treat it as supplemental, not standalone. Used well, remote signals like flood forecasts and self-enrollment data offer a fast-track list: individuals likely eligible for early support. Field teams and local partners are still essential to catch what the system misses and reach people who don’t have a phone at all.
GiveDirectly is ready to scale our use of AI-powered early cash aid
GiveDirectly continues to work with Google Research to scale the use of AI for anticipatory action to new countries and explore anticipatory cash transfers for other hazards, such as droughts. New research continues to prove the impact of large, early cash payments before climate disasters.
We have the data, the tools, and the evidence that acting before disaster strikes saves lives and livelihoods. Using AI to help send anticipatory cash should be a tool in the tool box for responding to climate shocks.
Note: this blog was updated in July 2026 to better reflect details and tradeoffs in this type of program design, as we have done elsewhere.
In the News
Appendix: Ward Selection
To identify which communities within Kogi were most vulnerable to floods, we used Google’s historical flood data, wealth indicators, and local insights. This data led us to Ibagi, a sub-district of Kogi. We then further pinpointed the most at-risk wards within Ibagi. Out of 41 wards we analysed, 6 were chosen (Onyedega, Unale, Iyano, Ayah, Ujeh, Ojila) based on three criteria:
- Wards with >1% population affected by Floods in last 5 years
- Wards with presence in the National Social Registry database (Nassco)
- Wards with high poverty rates
- The Global Flood Awareness System (GloFAS) is an operational global hydrological forecasting and monitoring system widely used around the world. ↩︎
- Common reasons why individuals may not have made it through the verification process include disclosing that they live outside the targeted area upon further investigation, not being known as a resident by community leaders, or were unable to be reached by our team after five phone calls and two in-person visits. ↩︎
