Helping the Worst-Off: When Hiring More Case Workers Beats Building Better AI

Government agencies rushing to deploy sophisticated prediction systems to identify vulnerable populations may be missing a simpler solution: hiring more staff. New research by incoming (Fall ’26) CDS and Courant Assistant Professor Juan Carlos Perdomo and collaborators reveals that expanding bureaucratic capacity often delivers greater improvements in helping those most in need than incrementally refining algorithmic models.

The findings, detailed in “The Value of Prediction in Identifying the Worst-Off,” challenge the current policy focus on perfecting prediction accuracy. Perdomo and his co-authors LMU Munich statistics PhD student Unai Fischer-Abaigar and LMU Munich Junior Professor Christoph Kern developed a framework to compare when prediction improvements outperform simply screening more people. This work won an Outstanding Paper Award at the International Conference on Machine Learning this year, a recognition given to the top six papers at the conference out of over 10,000 submitted.

“Prediction is a first and last mile effort,” Perdomo said. “You get significant improvement in social welfare by getting the system off the ground. If you have no prediction in place, having some basic predictor is useful, and at the end, if your predictor is near perfect, then you get substantial improvement from optimizing that final margin. But otherwise, you get much greater improvement by expanding bureaucratic capacity.”

The research emerged from Perdomo’s graduate work with the Wisconsin Department of Public Instruction, where officials were developing a dropout prediction system. Working with administrators made him question whether sophisticated prediction was worth the investment. “These are very busy people in under-resourced departments,” he said. “Is this really worth spending considerable time on? Or would creating a quick and simple model, then focusing on other interventions, be a better use of resources?”

To answer this question systematically, the researchers analyzed the “prediction-access ratio” (PAR) — a concept recently proposed in a prior paper by Perdomo — which quantifies how much more effective expanding screening capacity is compared to improving prediction accuracy. They tested their framework using administrative data on 274,515 German jobseekers over four decades, focusing on identifying those at risk of long-term unemployment.

The results were striking. For most real-world systems, which typically explain around 15–20% of variance in outcomes, expanding the number of people screened consistently outperformed prediction improvements. The researchers found that agencies need to screen about 25% more of the population than their target group size to account for imperfect predictions.

The mathematics revealed why this occurs. At very low prediction accuracy, any improvement helps enormously. At near-perfect accuracy, small gains can optimize an already excellent system. But in the middle range where most programs operate, the returns from better prediction are modest compared to simply reaching more people.

“When does investing in a prediction system make sense?” Perdomo asked. “By improving prediction, we improve the quality of our decisions, and therefore social welfare — reducing unemployment duration or increasing graduation rates. But in these problems, prediction is a means to an end.”

The research has immediate implications for cash transfer programs, job training initiatives, and social housing allocation. Rather than investing heavily in marginal prediction improvements, agencies might achieve better outcomes by hiring additional case workers or expanding their screening capacity.

The work also suggests that simple models may suffice for many applications. When the researchers compared a basic four-level decision tree to a sophisticated CatBoost model — a machine learning algorithm that combines many simple decision trees to make predictions — the simpler approach showed only a 5% decrease in predictive power while offering advantages in explainability and implementation costs.

Looking ahead, Perdomo sees this as the beginning of a new research direction that integrates prediction with resource allocation theory. “Computer science has historically focused on mechanism design and resource allocation questions that are fundamentally computational,” he said. “There’s now an opportunity to revisit these classical questions while incorporating statistics and learning theory.”

The findings provide a data-driven framework for budget directors and program administrators to make principled decisions about where to invest limited resources. In an era of AI enthusiasm, the research offers a sobering reminder that sometimes the most effective intervention is expanding human capacity rather than algorithmic sophistication.

By Stephen Thomas

Have feedback on our content? Help us improve our blog by completing our (super quick) survey.