Experimental Evidence of Agentic AI at Work

How to cite

Bojinov, Iavor. 2026. "Experimental Evidence of Agentic AI at Work." AEA RCT Registry. AEARCTR-0017902. https://doi.org/10.1257/rct.17902-1.0

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

Iavor Bojinov*

* Primary investigator

Registry ID
AEARCTR-0017902
DOI
https://doi.org/10.1257/rct.17902-1.0
First registered
Last updated
Status
In development
Trial dates
2026-02-17 – 2026-02-27
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/17902

Abstract

This study investigates how the development of AI agent design and governance choices shape quality and productivity outcomes in an enterprise setting. We have advised a large enterprise running a `lab-in-the-field' style experiment to measure how different agentic AI approaches affect work. The experiment compares the effectiveness of developing bespoke agents using local knowledge, using a centrally developed agent, and using standard, out-of-the-box generative AI tools (Microsoft Copilot M365). This research will provide insights into how organizations can effectively leverage agentic AI in complex knowledge workflows.

Experimental design

Participants will take part in a lab experiment. The lab experiment will include a pre- and post-survey designed to capture demographic information, generative AI usage and attitudes, and job attitudes. After completing the pre-survey, all participants will receive training on generative AI generally and an introduction to agentic AI. They will then complete experimental tasks using generative AI.

Design details

Randomization method
The company will perform a Bernoulli randomization using the Qualtrics randomization feature, where each participant will be randomly assigned to one of four treatments with equal probability (1/4).
Randomization unit
Individual participant.
Clusters
N/A
Observations
The company estimates that approximately 2,000 participants from a population of about 7,000 will take place. The primary factor for determining sample size is the participant opt-in rate of the employee population of roughly 7,000. Due to the dynamic nature of the enterprise environment, the employee population may be expanded.
Arms
The company will perform a Bernoulli randomization, where each participant will be randomly assigned to one of four treatments with equal probability (1/4).
Primary outcomes
Our primary outcome variable is quality of task output.
Secondary outcomes
In addition to quality, we will explore additional outcomes such as timing, delegation, and self-reported AI quality assessments.

Cited by

No citation data available for this registration yet.

Keywords