Generative AI and Productivity in the Global South

How to cite

Castro, Silvia; Mickeler, Maren; Ho, Hoa; Friess, Svenja; Griniute, Evelina. 2026. "Generative AI and Productivity in the Global South." AEA RCT Registry. AEARCTR-0018778. https://doi.org/10.1257/rct.18778-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Silvia Castro* , Maren Mickeler , Hoa Ho , Svenja Friess , Evelina Griniute

* Primary investigator

Registry ID
AEARCTR-0018778
DOI
https://doi.org/10.1257/rct.18778-1.0
First registered
Last updated
Status
In development
Trial dates
2026-07-01 – 2027-12-31
Countries
Private
JEL codes
G21, J24, O33, D22
Registry record
http://www.socialscienceregistry.org/trials/18778

Abstract

Generative AI (GenAI) tools are rapidly diffusing into professional workplaces, but causal evidence on their productivity effects outside high-income settings remains scarce. We study the introduction of a GenAI-enabled decision-support tool in a large financial-sector organization in Uganda, using a cluster-randomized field experiment in which knowledge workers are assigned to tool access or business-as-usual practice. Combining administrative records, tool-usage logs, and primary data, we estimate intention-to-treat effects on productivity, decision-making processes, and examine heterogeneity in tool adoption and usage to understand the mechanisms. The study provides novel experimental evidence on whether GenAI tools can extend productivity gains to lower-income labor markets, a central question for AI diffusion policy in the Global South.

Intervention

We introduce a customized GenAI tool that supports loan officers in their credit analysis of loan applications. The tool will be trained on the firm’s internal information base (policies, product databases, procedural manuals, previous product to client matches) and employees’ knowledge. Officers receiving the tool are also given training on how to use it, and the tool is available for routine use throughout the intervention period.

Experimental design

We conduct a branch-level randomized controlled trial at a commercial bank. Branches are matched into 40 pairs on pre-intervention characteristics, and within each pair one branch is randomly assigned to receive the intervention and one to serve as control (80 branches total, 50% treatment and 50% control). Randomization is conducted by computer at the branch level. Control branches continue under business as usual.

Design details

Randomization method
Pairwise matched randomization by computer
Randomization unit
Branch
Clusters
80 branches (40 pairs)
Observations
All loan officers employed in study branches at the time of randomization (except salary bankers); all active loans in the administrative records (except salary loans)
Arms
40 control and 40 treatment branches
Primary outcomes
Employee productivity, loan portfolio composition and performance
Sponsors
Name: Schmidt Sciences Url: https://www.schmidtsciences.org/ai-at-work/

Cited by

No citation data available for this registration yet.

Keywords