Creating AI Coworkers

How to cite

Perez-Truglia, Ricardo; Cullen, Zoe. 2026. "Creating AI Coworkers." AEA RCT Registry. AEARCTR-0019166. https://doi.org/10.1257/rct.19166-1.0

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

Ricardo Perez-Truglia* , Zoe Cullen

* Primary investigator

Registry ID
AEARCTR-0019166
DOI
https://doi.org/10.1257/rct.19166-1.0
First registered
Last updated
Status
On going
Trial dates
2026-07-13 – 2027-07-01
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/19166

Abstract

A large consulting firm is conducting an experiment with its employees and will share the resulting data with us so that we can study the treatment effects.

In brief, employees will have the opportunity to create and train their own AI coworkers: AI agents trained specifically on an employee’s knowledge. The main treatment arm randomizes incentives: half of the employees receive credit for the AI coworkers they create through recognition, visibility, and promotion opportunities. A secondary, cooperative treatment arm randomizes whether the company combines similar AI coworkers.

Intervention

A large consulting firm is conducting an experiment with its employees and will share the resulting data with us so that we can study the treatment effects.

In brief, employees will have the opportunity to create and train their own AI coworkers: AI agents trained specifically on an employee’s knowledge.

The main treatment arm provides incentives: employees receive credit for the AI coworkers they create through recognition, visibility, and promotion opportunities. A secondary, cooperative treatment arm randomizes whether the company combines similar AI coworkers.

Experimental design

The main treatment arm randomizes subjects into one of the following two groups with equal probability:

- Treatment: Employees can receive credit for the AI coworkers they create through recognition and career advancement internally, and IP they can take with them if they leave the company. They are informed about this credit in the same survey in which they decide whether to sign up.
- Control: Employees are not eligible for these incentives, nor are they informed that the incentives exist.

The hypothesis is that the provision of credit will be a strong enough incentive to persuade employees to create their own AI coworkers.

A secondary treatment arm randomizes employees into one of the following two groups with equal probability:

- Cooperation: Employees are told that the company may combine similar AI coworkers. For employees in the credit treatment group, this means that credit would be shared among creators in proportion to their contributions.
- Competition: Employees are told that the company will not combine similar AI coworkers, so they will be competing with their coworkers over who has the best AI coworker.

Like the conditional-decision outcome, this treatment arm is designed to study strategic complementarities.

The same survey in which employees decide whether to create AI coworkers also asks a few additional questions that can be used for heterogeneity analysis. The company will share administrative data that can be used for heterogeneity analysis too. We will have baseline measures of whether employees think their expertise is unique, whether they are planning to build and AI Coworker, and whether a similar AI Coworker has been or will be built. To test mechanisms we collect expressions of concerns and benefits to building an AI Coworker.

Design details

Randomization method
Randomization done in office by a computer
Randomization unit
The main treatment arms is randomized at the team level. The secondary treatment arm is randomized at the individual level.
Clusters
Subjects have to respond to a survey to make their decisions, but we do not know in advance what the response rate will be. We expect between 10-20%. The following numbers correspond to the total number of employees invited to the survey: around 5,000 employees divided in roughly 300 teams
Observations
It will depend on the response rate (with a maximum of 5,000 employees)
Arms
The main treatment arm will split the sample 50%-50%. The second treatment arm will split the sample 50%-50% too. They are cross-randomized.
Primary outcomes
The main outcome of interest is whether the employee signs up to create an AI coworker. This outcome in measured in two ways via a survey: as an unconditional decision and as a decision that can be contingent on the number of similar coworkers who also create an AI coworker. The conditional decisions will allow us to disentangle strategic complementarities. We may also be able to obtain a measure of intensity of interest based on whether employees who sign up go on to complete the intake form required to initiate the AI coworker creation process, as well as the time and effort they spend on that form.
Secondary outcomes
At the end of the survey, we ask employees questions about the oversight and control they would want to have in place for their AI Coworker, including who would ideally have access. Along the process to creating an AI Coworker, employees invest time in training and supplying data the AI Coworker knowledge.  Training and data supply will provide corroborating measures of our primary endpoints.

Cited by

No citation data available for this registration yet.

Keywords