AI Tool Use, Take-Up, and Learning

How to cite

Moore, Gideon; Bloom, Nicholas. 2026. "AI Tool Use, Take-Up, and Learning." AEA RCT Registry. AEARCTR-0018655. https://doi.org/10.1257/rct.18655-1.0

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

Gideon Moore* , Nicholas Bloom

* Primary investigator

Registry ID
AEARCTR-0018655
DOI
https://doi.org/10.1257/rct.18655-1.0
First registered
Last updated
Status
On going
Trial dates
2024-06-01 – 2026-12-31
Countries
Private
JEL codes
O, J, M
Secondary IDs
Stanford IRB: eProtocol 87086
Registry record
http://www.socialscienceregistry.org/trials/18655

Abstract

Despite the rapid rise of AI tools in knowledge work, there is limited evidence describing their effects on productivity or worker outcomes. In this project, we study the rollout of an AI tool suite to sales and support workers at a large technology firm over the course of several quarters in 2024-2025. We assess AI tool use not only on standard performance outcomes (e.g., sales, profits), but also on worker- and customer-focused outcomes: are gains driven by learning, where workers improve, or by selection, where less-productive workers attrit? When firms expand capacity with AI, which customers receive extra benefits? Findings will help firms, workers, and policymakers understand the labor market consequences of generative AI: who benefits, and through what mechanisms.

Experimental design

The experiment is a randomized controlled trial conducted by a large tech firm in the ordinary course of business. Sales and support employees were randomized into a treatment arm, which received access to a suite of generative AI tools (including real-time call-listening AI assistants that surface suggestions to workers), and a control arm, which continued to work without access to these tools. The experiment ran across several quarters in 2024-2025, after which the AI tool suite was rolled out company-wide. Stanford researchers do not interact with participants and observe only aggregate analyses produced by the large tech firm's causal inference team based on Stanford-directed analysis code.

Design details

Randomization method
Computer
Randomization unit
Employee and firm. Employees were randomly assigned to treatment or control, and all employees a given firm interacted with were in the same treatment group.
Clusters
All employees in this subsidiary of the large tech firm--thousands, along with their thousands of customers. Customers are randomized *with* employees; that is, although a customer may speak to different employees on different calls, all employees a customer speaks to will be within the same treatment arm.
Observations
Same as cluster
Arms
Determined internally by tech firm--will be disclosed to researchers following NDA
Primary outcomes
We are interested in both employee and customer outcomes of the experiment. Employee outcomes: * Pitch rate (replicating partner's prior analysis) * Win rate (replicating partner's prior analysis) * Take-Up * Satisfaction * Attrition * Subsequent learning * Heterogeneous effects by skill and tenure Customer outcomes: * Satisfaction * Future business interactions with large tech firm * Business success * Heterogeneous effects by firm size and tenure

Cited by

No citation data available for this registration yet.

Keywords