Automated Assignment and Bureaucrat Productivity: Evidence from a Field Experiment in Peru
How to cite
Ferreira, Alipio; Leon, Sandra Cermeno. 2026. "Automated Assignment and Bureaucrat Productivity: Evidence from a Field Experiment in Peru." AEA RCT Registry. AEARCTR-0018756. https://doi.org/10.1257/rct.18756-1.0
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
Governments increasingly use algorithmic tools to manage bureaucratic production, but these tools differ in whether they replace or support human discretion. We study this distinction in a pre-registered field experiment with Peru's national environmental enforcement agency. Analysts were randomly assigned to receive enforcement case lists generated either by managers or by an optimization algorithm. Independently, suggested completion dates were randomized at the case level. Replacing managerial assignment with algorithmic assignment reduced timely case completion by approximately 5 percentage points relative to a control mean of 40 percent, with no effects on downstream enforcement outcomes such as fines, appeal rates, or appeal outcomes. In contrast, suggested completion dates increased timely completion by approximately 4 percentage points, with effects driven by analysts with discretionary case assignment. We provide evidence that managers use discretion to match complex cases to experienced analysts, and the automated assignment reduces productivity by breaking this link.
Intervention
This paper evaluates two distinct automation treatments, implemented as a cross-randomized field experiment within OEFA's sanctioning unit. The first, which we call hard automation, replaces managerial discretion in case assignment: instead of team leaders allocating cases to analysts, an optimization algorithm generates the assignment list. The second, which we call soft automation, preserves the assignment decision but structures the information available to analysts: the algorithm produces suggested milestone dates for completing each procedural stage, displayed for a randomly selected subset of cases. These two treatments differ fundamentally in the degree to which they displace human judgment. Hard automation eliminates the team leader's role in matching cases to workers. Soft automation leaves that matching unchanged while providing analysts with scheduling information intended to help them organize their work.
Experimental design
The hard automation system generated case allocations using codified information on case characteristics and expected processing times. Relative to discretionary allocation, the algorithmic approach is transparent, generates a more even distribution of workload across workers, and is intended to reduce bias arising from poor managerial decisions. In contrast, discretionary assignment solves the same objective function as the algorithm but also incorporates soft information about worker-case fit, such as sector familiarity, writing quality, and reliability under time pressure. Both selection methods aim at maximizing output and preventing expiration risk by assigning cases under imperfect information about worker capacity and case complexity, but they differ in how they leverage that information. It is worth noting that the hard automation treatment is bundled: it simultaneously removes managerial judgment from case matching, and imposes workload equalization. The experiment therefore identifies the joint effect of these components rather than the effect of any single feature in isolation. However, we try to disentangle these two components in the mechanism analysis by providing evidence on how managers use discretion to match cases to workers.
Besides assigning cases to workers, the algorithm also suggested completion dates. This additional information could act as a nudge by helping analysts organize their work, regardless of how their cases were assigned. To assess the role of these dates, we randomized their display at the case level for 35% of control-group cases and for 50% of the cases handled by 22 randomly selected treated analysts. The remaining 34 treated analysts saw suggested dates for all of their cases. All suggested dates for conclusion lie before December 31, 2023, reflecting OEFA's intention to execute all the assigned cases in the same year.
Design details
Cited by
No citation data available for this registration yet.