When Algorithms Can Find the Poorest But Cash Still Misses Them

1. The problem

Low- and middle-income countries now spend an estimated hundreds of billions of dollars a year on social protection programs, yet a large share of that money never reaches households in extreme poverty. In many flagship cash transfer programs, 30–60% of benefits go to people who are not among the poorest, while 20–50% of the poorest are excluded entirely. That mis‑targeting blunts the impact of every dollar spent and keeps poverty rates higher than they need to be.

Traditional targeting methods are surprisingly crude. Many governments still use geographic targeting (sending benefits to all households in “poor” districts), community-based selection, or long paper “proxy means tests” that score assets like roof material or livestock to guess who is poor. These tools are slow to update, easy to game, and perform poorly when economies change quickly, such as during COVID‑19 or climate shocks.

The stakes are concrete. In Togo’s COVID emergency program, the goal was to reach informal workers suddenly out of income in a country where no national social registry existed. In rural East Africa, NGOs like GiveDirectly try to find households living on less than roughly the international extreme poverty line, but national surveys can be years out of date. Each percentage point of mistargeting represents tens of thousands of families missing out on cash proven to raise consumption, reduce food insecurity, and in some cases lower mortality.

2. What research shows

Over the past decade, researchers have shown that relatively simple machine‑learning models trained on household surveys and alternative data (like mobile phone metadata or satellite imagery) can match or beat traditional poverty targeting tools. In a widely cited multi‑country analysis, Blumenstock and co‑authors compared “algorithmic targeting” to geographic targeting and proxy means tests and found that, holding budgets constant, the ML approach could deliver more benefits to the poorest for large-scale programs. While full confusion matrices vary by setting, simulations frequently show 5–15 percentage‑point reductions in exclusion errors compared with geographic targeting at similar cost.

A flagship real‑world case comes from Togo’s Novissi emergency cash program during COVID‑19. Researchers trained models on survey and mobile phone data to predict household consumption, then used those scores to identify poor mobile subscribers in urban areas. When they compared targeting regimes ex post, the ML phone‑data approach reduced exclusion errors by 4–21 percentage points relative to realistic geographic options considered by the government. Relative to a hypothetical, perfectly updated social registry, ML targeting had 9–35 percentage‑point higher exclusion errors, showing it was not perfect but clearly better than the status quo.

Other studies show similar patterns. Work using call-detail records and satellite imagery in multiple African countries has found that ML models explain 50–70% of the variation in consumption or asset indices, often outperforming linear models or simple wealth scores. Simulations suggest that when budgets are fixed, these methods can improve “benefits to the poorest” by 10–30% compared with geographic targeting, especially in highly unequal regions.

3. What the real world shows

The most compelling evidence comes from prospective implementations where ML targeting actually shaped who got money. In Togo’s Novissi program, the government and collaborators used the ML phone‑data model to disburse roughly $10 million in COVID relief to informal workers between 2020 and 2021. The program reached more than 57,000 poor households in urban areas, and evaluation showed it materially reduced exclusion errors versus the geographic targeting the government had initially considered.

NGO cash transfer programs add outcome evidence on what better targeting buys you. A randomized study in Kenya found that a one‑time transfer of about $1,000 to poor rural households reduced infant deaths by 48% relative to control communities over the study period. GiveDirectly and partners have delivered over $58 million in basic income and lump‑sum transfers to more than 56,000 people across several African countries, consistently documenting increases in consumption, asset investment, and psychological well‑being. While these trials often use randomization rather than algorithmic targeting, they demonstrate that every correctly targeted household can generate large health and economic gains.

Recent work on anticipatory humanitarian action reinforces the point. A machine‑learning model used by humanitarian agencies to forecast displacement across 26 countries achieved a median absolute error of about 11% between forecasted and actual displacement numbers, using that to plan responses 3–4 months ahead. Forecast-based financing pilots by the Red Cross and UN agencies have used similar models to pre‑position cash and supplies before floods or cold waves, reducing later emergency costs and losses, although robust meta‑estimates of cost savings are still emerging.

A growing set of systematic and scoping reviews synthesize these findings. Recent reviews of data‑driven social protection targeting and humanitarian anticipatory action (2020–2025) conclude that ML methods generally outperform simple geographic or categorical targeting in retrospective tests and that early field deployments show promising improvements in inclusion of the poorest, but evidence on long‑run political and social impacts remains limited.

4. The implementation gap

Despite the performance gains, very few national social protection systems have moved from pilots to routine ML‑based targeting. Surveys of low‑ and middle‑income countries find that most still rely on categorical eligibility (such as “all children under five”) and community‑based targeting, with about one‑third using proxy means tests and only a small minority experimenting with ML or big‑data approaches. A reasonable estimate is that fewer than 5% of large national programs currently use ML-based targeting as a core mechanism.

Why the hesitation? One barrier is data. High-performing models like the Togo phone‑data system depend on detailed call‑detail records or mobile money data, which governments may not legally or politically be able to access. Even where regulators permit it, telecom operators may be reluctant to share granular data, and there are real privacy concerns about inferring poverty from digital traces. Many countries also lack recent, high‑quality household surveys to train initial models, especially for informal settlements and marginalized groups.

Another barrier is trust and transparency. Proxy means tests and community meetings may be imperfect, but officials and citizens can at least see the questions and criteria. By contrast, even relatively simple ML models can seem like black boxes, especially when they ingest opaque features from phone metadata or satellite imagery. Civil society groups worry that algorithmic scores could encode or amplify existing discrimination, for example by underserving people without phones or women whose phone usage looks different from men’s.

There are also operational and political frictions. Social ministries often lack in‑house data science capacity to build, monitor, and update models, and outsourcing to vendors raises concerns about lock‑in and accountability. Politically, leaders may prefer geographically broad or categorical programs that spread benefits visibly, rather than tightly targeted schemes that optimize for poverty reduction but leave many non‑poor voters out. Implementation pilots have also exposed mundane but serious issues like difficulties enrolling people identified by phone data in remote regions, mismatches between telecom data and civil registries, and disputes when neighbors see some households get cash while others—whose poverty is less “visible”—do not.

5. Where it actually works

Some contexts have managed to bridge the gap. Togo’s Novissi program is now the canonical example: a government with no national social registry partnered rapidly with researchers and a mobile operator to build an ML‑based targeting system, backed it with regulatory approvals, and rolled it out nationwide in an emergency. Key enablers included strong political backing, a single telecom with large market share, and clear rules limiting data use to the emergency relief program.

NGOs and multilaterals have also made progress in controlled pilots. GiveDirectly’s work using satellite imagery to target poor villages in Kenya and Uganda, for example, has shown that automated village selection can approximate survey-based methods at much lower cost, freeing more money for transfers. Humanitarian agencies using displacement-forecast models in the Sahel have embedded data teams inside operations units and built governance processes that allow field staff to override model recommendations when local knowledge contradicts the forecast.

6. The opportunity

Better targeting will not eliminate poverty on its own, but the evidence suggests that using data‑driven methods in large cash programs could redirect millions of dollars per country each year toward households in deepest need.

To move from pilots to practice, the most promising steps are:

Build public data infrastructures (regular household surveys, safe data‑sharing agreements with telecoms) that make ML targeting feasible and auditable.
Invest in in‑house analytic capacity within social ministries so models can be owned, updated, and scrutinized locally rather than treated as vendor magic.
Design hybrid systems that combine ML scores with community input and simple appeal processes to reduce bias and increase perceived fairness.
Strengthen privacy and governance frameworks so citizens know how their data is used and have recourse if systems go wrong.
Tie pilots to clear impact metrics (like exclusion error, share of transfers going to the poorest, and downstream outcomes) and publish results, building a norm that “what works” gets scaled.

Blumenstock, J., et al. “Scalable Targeting of Social Protection: When Do Algorithms Outperform Traditional Methods?” NBER Working Paper 33919, 2023.
Aiken, E., et al. “Machine learning and phone data can improve targeting of humanitarian assistance.” Nature 603, 864–870, 2022.
World Bank. “Truncated Early Stopping for Proxy Means Testing.” Policy Research Working Paper, 2020.
DataFriendlySpace. “The Role of Predictive Analytics in Humanitarian Operations.” 2025.
Camara, B., et al. “Pushing the boundaries of anticipatory action using machine learning.” Data & Policy, 2025.
GiveWell. “GiveDirectly’s Cash for Poverty Relief Program.” 2022.
J-PAL. “Giving directly to support poor households.” Case Study, 2025.
GiveDirectly. “Our Research.” Research summaries page, accessed 2025 (includes Kenya infant mortality RCT).
Innovations for Poverty Action. “GiveDirectly Case Study – Goldilocks Toolkit.” 2016.
Blumenstock, J. “Scalable Targeting of Social Protection: When Do Algorithms Outperform Traditional Methods?” Working paper version, 2023.