An Algorithm Chose Where 2,000 Refugees Would Live. Employment Rose 10% for CHF 50 Each

Between January 2020 and June 2023, Switzerland let an algorithm suggest where about 2,000 refugee cases would live. Neither the officers making the decision nor the refugees knew whether the suggestion came from the algorithm or from something resembling the old procedure.

Three years later, the algorithm's group had spent about 10% more of their months in work. The cost of running it was estimated at CHF 50 per adult. A paper posted to arXiv on September 28, 2026 by researchers at Stanford's and ETH Zurich's Immigration Policy Lab, UC Berkeley, Dartmouth and Harvard Business School reports what the authors describe as the first randomized trial of algorithmic refugee matching.

It is a rare case of AI in the labor market that adds jobs rather than removes them. It is also more fragile than the headline suggests.

A decision the government was making anyway

[Fact] Switzerland assigns asylum seekers to its 26 cantons using proportional quotas and national-origin balancing rules, without interviews. Placement officers have little information about where a particular person is likely to find work.

[Fact] The Swiss State Secretariat for Migration (SEM) ran a tool called GeoMatch. It uses administrative records and machine learning to predict each case's employment chances in every canton, then picks the canton that maximizes predicted employment across all cases, within the quotas. Officers kept final authority.

[Fact] The design detail that matters: the treatment and control arms each had their own identical copy of the canton and origin quotas, so they never competed for the same slots. Any gain has to come from sending people to cantons that suit them better, not from steering everyone toward Zurich's labor market.

What changed for refugees

[Fact] The pre-registered primary outcome was the share of months employed during the first three years. The control group averaged 22.3%. The algorithm's group scored +2.2 percentage points higher, about 10% in relative terms (95% confidence interval +0.05 to +4.33, p = 0.045).

[Fact] The effect grew with time. At 36 months, the share of cases with any adult employed was +5.2 points higher on the pooled sample, about 11%. Measured per person, the three-year employment rate rose 4.6 points on a control base of 44.8%.

[Fact] Adherence was not an issue. Officers followed the displayed recommendation 97% of the time.

The COVID problem hiding in the average

Here is where the pooled number needs care.

[Fact] GeoMatch was trained on pre-pandemic data, and the trial launched in January 2020. For the 2020 and 2021 placement cohorts, the estimated effects are statistically indistinguishable from zero. The gains come from 2022 and 2023, where the effect on the primary outcome was +3.9 points (about 17%).

[Fact] The authors explain why. Refugees who found work were concentrated in hospitality, other face-to-face service jobs, cleaning and construction, exactly the sectors lockdowns shut. A model that had learned which cantons absorb new arrivals into restaurants was predicting a world that had stopped existing.

[Fact] The authors are careful with this: the cohort comparison was not pre-registered. The treatment-by-era difference has p = 0.039 with covariates and p = 0.070 without.

[Claim] That split is the most useful thing in the paper for anyone deploying predictive tools. The same algorithm, run the same way, produced nothing for two years and a solid gain for the next two. The difference was whether the world still looked like the training data.

The cost comparison, in plain numbers

[Fact] The authors value each extra month of employment at CHF 1,950 in fiscal terms (lower support payments plus social contributions and source taxes). The pooled effect equals 0.70 extra employment-months per adult over three years, roughly CHF 1,300 net per adult after a CHF 50 operating cost. That is a benefit-cost ratio of about 27 to 1, counting only the government's side of the ledger.

[Estimate] 0.70 months over 36 is about three extra weeks of work per adult. Small per person, but the tool costs almost nothing per person.

[Fact] The paper compares it to heavier programs. Denmark's 1999 reform added about 430 hours of language training and raised full-time-equivalent employment by 4.2 points on average over 18 years. Germany's 600-hour integration courses raised employment by about 5 points after one year. Randomized trials in Sweden and Italy found gains of 15 and 10 points, at roughly CHF 2,400 and CHF 3,000 per participant.

[Estimate] Put side by side, GeoMatch's per-person gain is two to three times smaller than the Swedish and Italian programs, and 48 to 60 times cheaper per participant. That is the actual argument. It is not a better program. It is a nearly free improvement to a decision that has to be made regardless.

Reasons to hold the result loosely

• [Fact] The authors built the tool they are evaluating. GeoMatch comes from the Immigration Policy Lab. They declare no competing interests and state they received no payments from SEM. Funding came from Stanford HAI, Schmidt Futures, Google.org, the Rockefeller Foundation, Stanford Impact Labs, the Charles Koch Foundation, Open Society Foundations and the European Research Council.
• [Fact] The lower confidence bound on the pooled effect is +0.05 points, barely above zero.
• [Fact] The sample is specific: mean age 27.9, 16% women, and 54% from Afghanistan, 26% from Turkey and 9% from Syria.
• [Fact] The outcome is whether someone worked, not what they earned or whether the job was a good one.
• [Claim] With 97% adherence, "human officers retain final authority" means little in practice. That is fine when the tool works. It is a concern in a 2020-type shock, when the model is wrong and nobody overrides it.

What this means for work

Most of what we cover is AI changing tasks inside an existing job. This one is different: AI changes which job a person can reach, by changing where they start.

• For refugees and the people who help them, the first placement shapes the next decade. The paper cites European evidence that refugees remain about 12% less likely to be employed than comparable migrants for 10 to 15 years.
• For caseworkers, placement officers and HR matching roles, the lesson is not replacement. The officer still decides. But when a recommendation is followed 97% of the time, the skill that matters shifts to knowing when the model's world has changed.
• For the entry jobs themselves, the jobs these refugees reached are among the least exposed to generative AI: cooking, serving, cleaning, building. Better matching raises employment in work that is hard to automate.

For occupation-level data, see our pages on cooks, waiters and waitresses, janitors and cleaners, housekeepers, construction laborers and human resources specialists.

Sources

• Kirk Bansak, Jens Hainmueller, Dominik Hangartner, Jeremy Ferwerda, Elisabeth Paulson, Angie Delevoye, Nicholas Adams-Cohen, Ashwin Ramaswami, Selina Kurer, Joelle Pianzola and Michael Hotard, "AI-based matching improves refugee employment in a double-blind randomized trial," arXiv:2609.35448, September 2026: https://arxiv.org/abs/2609.35448

About this analysis

This article was written with AI assistance. All figures marked [Fact] were read from the full text of arXiv:2609.35448v1 (36 pages of main text plus 22 pages of supplementary information), not from the abstract alone. The paper is a preprint and has not yet been peer reviewed. Figures marked [Estimate] (the three-weeks conversion and the 48-to-60-times cost comparison) are our own arithmetic from the paper's reported numbers. The statement that the refugees' entry jobs have low generative-AI exposure is our characterization, not the authors'. No tables or figures from the paper are reproduced.