AI Matching Raises Refugee Employment in Swiss Trial

Where a refugee is first placed can shape access to jobs, local networks, language environments, and public services for years. Yet placement decisions must be made quickly and usually with incomplete evidence about which destination suits a particular case. This study tests whether algorithmic decision support can improve that match in a live government allocation system rather than in a retrospective simulation. In a Swiss double-blind randomized trial covering roughly 2,000 refugee cases placed from January 2020 through June 2023, the employment-optimized recommendation produced higher subsequent employment than a recommendation designed to approximate existing practice.

Core Contribution

The central contribution is field evidence that a matching algorithm can improve a high-stakes public allocation decision while preserving human authority over the final placement. The intervention did not simply direct more people to cantons with better labor markets. The treatment and control arms operated with identical but separate quotas by canton and origin group, so the estimated difference isolates better pairing of individual refugees with destinations under the same aggregate allocation constraints.

That design matters. Many claims for predictive public-sector systems rest on offline accuracy or on reallocations that would be difficult to implement politically. Here, the relevant question is narrower and more practical: given the placements a system is permitted to make, does a recommendation improve outcomes? The answer is positive on the pre-registered primary outcome, though the size of the gain is incremental rather than transformational.

Technical Approach

The authors' system combines administrative data, machine learning, and constrained optimization. As refugee cases arrive, it recommends a canton intended to maximize employment prospects while satisfying the operational quota structure. Placement officers retain final decision authority, and both officers and refugees were blinded to whether an algorithmically optimized or control recommendation had been assigned.

The randomized comparison is especially consequential because matching systems are exposed to a basic selection problem: refugees sent to a given canton under ordinary procedures may differ systematically from those sent elsewhere. Random assignment of cases to recommendation arms addresses that problem. Separate quota pools also prevent treatment effects from being attributed to a changed canton mix. The design therefore evaluates decision support under a constraint that resembles the actual administrative setting, rather than an unconstrained assignment exercise.

The trial began immediately before COVID-19 altered labor-market conditions. The paper consequently reports pooled estimates over the 2020--2023 placement cohorts and distinguishes the later 2022--2023 cohorts. That temporal split is not a cosmetic detail: the estimated effect is larger after the pandemic-era disruption, indicating that matching value depends on the labor-market environment in which placements occur.

Results and Analysis

For the primary outcome, the share of months employed in the first three years after placement, algorithmic recommendation increased the pooled intention-to-treat estimate by 2.2 percentage points relative to a 22.3% control mean. This is approximately a 10% relative increase, with a 95% confidence interval from 0.05 to 4.33 percentage points. The interval narrowly excludes zero, so the pooled evidence supports a positive effect but also leaves uncertainty about the exact magnitude.

The result is stronger for refugees in the 2022--2023 cohorts: employment exposure rose by 3.9 percentage points, or about 17% relative to the same 22.3% reference level, with a 95% confidence interval of 1.11 to 6.68 percentage points. Effects also accumulated rather than appearing only at placement. At 36 months, the employment-rate ITT estimate was 5.2 percentage points higher, an approximately 11% increase, with a 95% confidence interval of 1.10 to 9.25 percentage points.

These are meaningful effects for an intervention that changes a recommendation rather than delivering intensive services. The authors compare the 36-month gain to the employment benefit associated with hundreds of hours of intensive language training. Still, the comparison should not obscure the mechanism: the trial demonstrates a gain from assignment quality within Swiss quota rules, not a substitute for language instruction or evidence that an algorithm can resolve broader barriers to labor-market integration.

Limits for Deployment

The evidence is strong for this policy setting, but its scope is specific. The outcome is employment over the first three years, not earnings, job quality, housing stability, social integration, or long-run residence outcomes. The recommendation system was evaluated in Switzerland under its own canton and origin-group quotas; other countries may have different labor markets, administrative data, legal rules, and destination capacity. Finally, COVID-era changes coincide with the trial period, and the larger later-cohort estimate shows that effects may vary with labor-market conditions. The study supports constrained algorithmic matching as a useful administrative tool, not a universal rule for refugee placement.