What Happens to Welfare Enrollment When Human Caseworkers Are Replaced by Automated Systems?

Wu and Meyer (2025) examined how automating welfare caseworker services affected enrollment in SNAP, TANF, and Medicaid programs in Indiana. Their primary research question asked what happens to welfare participation and program targeting when automated systems replace face-to-face caseworker assistance. They analyzed administrative records covering nearly 3 million welfare recipients linked to IRS income data during Indiana’s phased rollout of an IBM automation system between 2007 and 2009. They found that automation reduced SNAP enrollment by 15%, TANF enrollment by 24%, and Medicaid enrollment by 4%. Their results also showed that more needy recipients were especially likely to lose benefits during recertification under the automated system.

Why This Article Was Selected for The Policy Scientist

This article addresses a policy issue that has become increasingly important as governments rely more heavily on automated systems to administer public services. The study is especially timely because state and federal agencies continue to replace face-to-face interactions with online platforms, call centers, algorithmic reviews, and automated eligibility systems in welfare, unemployment insurance, and health programs. Wu and Meyer make an important contribution by examining not simply whether automation changes enrollment, but also which types of recipients are screened out during application and recertification. Their findings help clarify broader debates in the administrative burden literature associated with scholars such as Herd, Moynihan, and Finkelstein. The article is particularly valuable because it analyzes multiple programs simultaneously rather than focusing on a single policy domain.

The data quality is exceptionally strong. The authors rely on administrative records covering nearly 3 million recipients linked to IRS income data, which substantially reduces measurement error compared to survey-based studies. The phased county-level rollout of Indiana’s IBM automation system creates a credible natural experiment that supports causal inference through a difference-in-differences design. While not a randomized controlled trial, the research design is substantially stronger than conventional multivariate regression approaches commonly used in policy research. The longitudinal structure of the data also allows the authors to distinguish between entry and recertification effects, which strengthens the analysis considerably. Although Indiana’s experience involved a particularly flawed implementation, the findings are likely generalizable to other jurisdictions pursuing automated welfare administration, especially given the widespread post-pandemic expansion of remote eligibility systems.

Full Citation and Link to Article

Wu, D., & Meyer, B. D. (forthcoming). Certification and recertification in welfare programs: What happens when automation goes wrong? American Economic Journal: Economic Policy. https://doi.org/10.1257/pol.20250309

Central Research Question

Wu and Meyer examine how administrative burdens generated by automated welfare systems affect enrollment, recertification, and targeting across major public assistance programs. Specifically, they ask what happens when face-to-face caseworker assistance is replaced with automated systems that rely heavily on call centers, online processing, and standardized eligibility procedures. The article focuses on Indiana’s large-scale welfare automation initiative implemented through IBM between 2007 and 2009. The central question is not merely whether enrollment falls after automation, but which populations are disproportionately screened out during different stages of the application process. The authors are especially interested in understanding whether automation primarily deters less needy applicants at entry or instead removes more vulnerable recipients during recertification. By analyzing SNAP, TANF, and Medicaid simultaneously, the paper also investigates why administrative burdens produce different effects across welfare programs with different recertification structures and eligibility rules.

Previous Literature

The article builds on a substantial literature concerning administrative burden, welfare participation, and program take-up. Earlier research by scholars such as Currie, Herd, Moynihan, and Finkelstein demonstrated that procedural barriers can meaningfully reduce participation in social programs even among eligible populations. Some studies suggest that enrollment burdens improve targeting by discouraging less needy households with higher opportunity costs of time, while other studies conclude that burdens disproportionately harm vulnerable populations with limited administrative capacity or unstable living conditions.

Wu and Meyer attempt to reconcile these competing findings by emphasizing the importance of enrollment stage. They argue that prior studies often examined only one program or one portion of the application process, thereby obscuring important differences between initial application and recertification. Their theoretical framework suggests that the composition of applicants changes dynamically over time. Less needy individuals may be screened out during initial application because they face higher opportunity costs, while more vulnerable recipients may fail to complete recertification once already enrolled. This distinction allows the authors to explain why prior empirical studies sometimes reached contradictory conclusions despite examining similar administrative burdens.

The article also contributes to a growing literature on the automation of public services. Previous work frequently examined the benefits of technological modernization, including online applications and automatic enrollment systems. However, fewer studies have evaluated the unintended consequences of replacing human caseworkers with centralized automated systems. The authors situate Indiana’s experience within a broader trend toward remote administration in welfare programs, unemployment insurance systems, and public health programs. The article therefore extends the administrative burden literature into the context of digital governance and automated service delivery.

Data

The study relies on unusually comprehensive administrative data. The authors analyze records covering nearly 3 million recipients participating in SNAP, TANF, and Medicaid in Indiana. These administrative records include monthly information on benefit receipt, enrollment status, demographic characteristics, and county of residence. The data span the period surrounding Indiana’s phased rollout of the IBM automation system.

A major strength of the dataset is its linkage to Internal Revenue Service microdata. Through these linked records, the authors obtain detailed measures of income, wages, and economic well-being. These linked administrative datasets substantially reduce the measurement error common in survey-based welfare research. Because the data follow recipients longitudinally over time, the authors can separately analyze entry into programs, continued participation, and exits from enrollment.

The authors also supplement the microdata with county-level economic and demographic indicators, including unemployment rates, industry employment composition, poverty rates, and population characteristics. These additional measures help evaluate whether treated and untreated counties experienced similar economic trends before the policy intervention. The data reveal that treated and untreated counties displayed remarkably parallel trends in welfare participation and unemployment prior to automation, strengthening the credibility of the research design.

The overall quality of the dataset is exceptionally high for social policy research. Administrative records provide near-universal coverage of program participants, while the linkage to IRS records permits unusually detailed analyses of recipient need and targeting. Few studies in the welfare literature combine such extensive administrative and tax data at this scale.

Methods

The authors employ a generalized difference-in-differences design based on the staggered rollout of Indiana’s automation system across counties. Because IBM’s system was introduced in different regions at different times before the rollout was halted, the authors can compare changes in welfare participation between treated and untreated counties over time.

The empirical strategy approximates a natural experiment. The authors demonstrate that treated and untreated counties followed similar pre-treatment trends in program enrollment and economic conditions, supporting the parallel trends assumption underlying causal inference. Their analysis incorporates county fixed effects and time fixed effects to account for unobserved local differences and statewide shocks, including the Great Recession.

Importantly, the study goes beyond conventional enrollment analyses by decomposing effects into entry and exit margins. This distinction allows the authors to identify whether automation primarily discouraged new applicants or instead caused existing recipients to lose benefits during recertification. The authors also estimate heterogeneous effects across programs and recipient characteristics.

In addition to the empirical analysis, the paper develops a dynamic theoretical model of welfare participation. The model conceptualizes enrollment as a recurring process involving both application and recertification decisions. This framework helps explain why administrative burdens can generate opposite targeting effects at different stages of program participation.

The methodological approach is notably strong. Although the study is not a randomized controlled trial, the staggered implementation of the policy creates a credible quasi-experimental setting that supports causal interpretation. The use of longitudinal administrative records further strengthens identification by allowing precise tracking of recipient behavior over time. Relative to standard multivariate regression approaches commonly used in public policy research, the design provides substantially stronger evidence regarding causal effects.

Findings/Size Effects

The results indicate that Indiana’s automation system substantially reduced welfare participation across all three major programs examined. One year after implementation, SNAP enrollment declined by approximately 15%, TANF enrollment declined by 24%, and Medicaid enrollment declined by 4%. These effects were statistically significant and emerged shortly after automation began.

The largest declines occurred in TANF, which the authors attribute partly to the program’s more frequent recertification requirements and greater documentation burdens. Medicaid experienced the smallest short-run decline because recipients recertified less frequently. Nevertheless, Medicaid enrollment remained approximately 2% lower in treated counties even four years after the automated system was abandoned, suggesting durable long-term effects.

The decomposition of entry and exit margins produces some of the paper’s most important findings. Automation reduced entry into all programs, but it also increased exits during recertification, particularly in TANF. The authors find that less needy individuals were disproportionately screened out during initial application, while more vulnerable recipients were more likely to lose benefits during recertification. Recipients exiting during recertification tended to have lower incomes, greater disability levels, and fewer economic resources than those screened out during initial application.

The authors argue that these results reconcile prior disagreements in the literature regarding administrative burden and targeting. Administrative burdens can simultaneously discourage less needy applicants at entry while removing highly vulnerable recipients already participating in programs. The effect depends critically on where individuals are located in the enrollment process.

Conclusion

Wu and Meyer conclude that administrative burdens generated through flawed automation systems can substantially alter both welfare enrollment levels and the composition of recipients receiving benefits. Their findings demonstrate that replacing personalized caseworker assistance with centralized automated systems may generate significant participation declines even when eligibility rules remain unchanged.

The article’s broader contribution lies in demonstrating that administrative burden is not a static phenomenon. The effects of burdens vary across programs, populations, and stages of participation. By distinguishing between application and recertification, the authors provide a more dynamic framework for understanding how bureaucratic systems shape access to public benefits.

The study also contributes to contemporary debates surrounding digital governance and automated public administration. As governments increasingly rely on online platforms, algorithmic processing, and remote service delivery, the Indiana experience offers evidence regarding the potential consequences of poorly implemented automation systems. Although Indiana represented an especially problematic rollout, the broader mechanisms identified in the article likely extend to many other jurisdictions that have shifted toward automated welfare administration.

Overall, the paper represents a substantial empirical and theoretical contribution to the literature on welfare participation, administrative burden, and public sector automation. Its combination of high-quality administrative data, strong causal inference methods, and dynamic theoretical modeling makes it one of the more rigorous recent studies examining the interaction between technology and social policy administration.