Humans + AI in Commercial Banking

How to cite

Su, Yang; Zhu, Wu; Yu, Jiaheng; Zhang, Leifu. 2025. "Humans + AI in Commercial Banking." AEA RCT Registry. AEARCTR-0016412. https://doi.org/10.1257/rct.16412-1.0

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

Yang Su* , Wu Zhu , Jiaheng Yu , Leifu Zhang

* Primary investigator

Registry ID
AEARCTR-0016412
DOI
https://doi.org/10.1257/rct.16412-1.0
First registered
Last updated
Status
On going
Trial dates
2025-07-01 – 2027-09-30
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/16412

Abstract

The rapid development of AI has greatly affected the workplace, with even greater potential to be unleashed in the short future. While in simple tasks AI may completely replace humans, in more complicated tasks the approach is more likely to be "AI + Humans", where humans oversee the actions of AI and humans make the final decisions. This approach is based on the assumption that such "AI + Humans" approach can out-perform Humans alone or AI alone. We examine the performance of "AI + Humans" for the credit screening task in the commercial banking sector using randomized controlled trials.
We will first develop an AI assistant by training the popular general AI models using the bank historical lending data. We will then ramdonly open this AI assistant to human loan officers so that these human loan officers can seek advice from the AI assistant. We then examine how this AI-assisted humans perform relative to the AI assistant alone or other human loan officers alone. In particular, in scenarios where AI make more (less) accurate credit risk assessment, whether and when humans follow (ignore) AI's advices? What are the policy implications that can lead to better human uses of AI assistant?

Intervention

We will develop an AI assistant by training the popular general AI models using the bank historical lending data. The intervention would be: a random group of human loan officers will be allowed to use this AI assistant while others cannot.

Experimental design

We will open the AI assistant to a subset of human officers. We shall require them to upload relevant loan application information for the AI assistant to process. We will record the human loan officers' assessment, the AI assessment, and the final action taken by the human officers.

Design details

Randomization method
We will assign the AI assistants to human loan officers with staff ID that ends with an odd number.
Randomization unit
Individual.
Clusters
200 loan officers.
Observations
5000 loan applications.
Arms
100 loan officers control, 50 loan officers with access to AI Assistant Version 1, 50 loan officers with access to AI Assistant Version 2.
Primary outcomes
The loan application approval rate and the performance of approved loans.

Cited by

No citation data available for this registration yet.

Keywords