Annie MOORE Decides Where Refugees Are Most Likely to Thrive

When a refugee family arrives in the United States, one decision shapes almost everything that follows: where they are placed. The wrong city can mean isolation, unemployment, and a second displacement. The right one can mean a job within weeks and a community that helps a family rebuild. A software system called Annie MOORE now makes that decision dramatically better. Built by researchers at the Universities of Oxford and Lund and Worcester Polytechnic Institute together with the resettlement agency HIAS, it uses machine learning and integer optimization to place each family where they are most likely to find work. HIAS analysis found that if the system had been used for all of its 2017 placements, more than 20 percent additional refugees would have found jobs within their first 90 days.

The tool is named after Annie Moore, the 17 year old Irish immigrant who in 1892 became the first person processed at Ellis Island. The modern MOORE stands for Matching and Outcome Optimization for Refugee Empowerment. Where the original Annie Moore stepped off a ship into an unknown country, her namesake tries to make sure today's arrivals step into a place that gives them a genuine chance.

Placement Decides Outcomes

The United States gives refugees little time to settle. Support periods are short, and families are expected to find work and become self sufficient quickly. Yet for years, placement decisions were made largely on informal criteria: which affiliate office had capacity, where people speaking the same language lived, where housing was available.

Those criteria matter, but they ignore something researchers consider decisive. The same family can have very different employment odds in different cities depending on local labor markets, existing community networks, and the family's own characteristics. Research published in Operations Research showed that initial placement profoundly affects long term outcomes, yet resettlement decisions had rarely been optimized at all.

The numbers from HIAS make the stakes concrete. Of the 496 employable refugees the agency resettled in 2017, 159 found some form of employment within 90 days. When researchers ran that same caseload through Annie MOORE in simulations, more than 200 would have found jobs. Same people, same resources, same year. The only difference was the matching.

How the Matching Works

Annie MOORE works in two layers. First, a machine learning model trained on years of past arrival data estimates how likely a given refugee or family is to find employment in each of the roughly 200 communities in the HIAS network. An Arabic speaking family of five from Syria might be served by many affiliates, but the model can distinguish between a city where their profile historically leads to work and one where it does not.

Second, an integer optimization engine solves the assignment problem across all current cases at once. It balances employment likelihood against real constraints: affiliate capacity, family indivisibility, funding rules, and language support. The output is a recommended placement for every family in the queue, not just the single best city for each one in isolation.

Caseworkers stay in control. The software recommends, and staff can override any suggestion with a manual placement. In practice HIAS found the opposite of replacement. Karen Monken, the agency's Associate Director for Pre arrival, reported that her team could stop spending time on routine matching and focus on genuinely difficult cases, such as those involving significant medical conditions.

From Pilot to Daily Use to Annie 2.0

HIAS began testing the system in 2017 and put it into production use for placements in May 2018. Oxford reports that Annie has matched well over 1,100 refugees resettled by the agency, and the research team estimates the software obtained an employment boost above 30 percent in its early period, moving measured employment rates from roughly 30 percent to around 40 percent for the cases it influenced. The system has also measurably increased the agency's operational capacity, letting staff process more cases without more people.

In 2024 the team published a major upgrade described in Communications of the ACM. The new dynamic placement algorithm accounts for uncertainty about future arrivals, batching of cases, and families that cannot be split apart. In tests on historical HIAS data, the algorithms achieved over 98 percent of the hindsight optimal employment level in every year studied, while the simpler greedy approach used before typically reached only about 90 percent. The upgraded logic was folded into the software the agency continues to use.

The research also spread. A related system called GeoMatch, developed by a overlapping research community, has been piloted by the Swiss State Secretariat for Migration since 2020. During the response to the war in Ukraine, HIAS used a companion tool called RUTH to match Ukrainian refugees with American volunteer hosts. The idea that placement is a computational problem worth solving well has become a small but growing field.

What This Says About Algorithms and Vulnerable People

Using an algorithm to decide where refugee families live sounds cold, and the researchers have been careful about the ethics. Every recommendation can be overridden by a human. The optimization objective, early employment, was chosen because self sufficiency is the explicit goal of the US resettlement program itself. Academic critics have continued to probe the approach, including a 2025 reevaluation of location matching algorithms, which is exactly the kind of scrutiny a system like this deserves.

The deeper lesson is about wasted potential. For decades, resettlement agencies made high stakes placement decisions with spreadsheets, intuition, and whatever capacity happened to be open. The data to do better already existed in agency records. It took researchers willing to sit inside a humanitarian bureaucracy and turn that data into a working tool. Over 20 percent more jobs, from a better match. That is what AI for good can look like when the problem is not diagnosing a disease or spotting a wildfire, but simply asking where a family should be allowed to begin again.