Moving Laggards: Nudges and Training for Enterprise AI Adoption

Generative AI is being deployed across the corporate world at unusual speed. The early causal evidence on its productivity impact is encouraging but it comes from small samples in specific tasks. We still know little about how workers in large enterprises adopt these tools when they are made available and even less about which interventions move adoption among the employees who feel reluctant or overwhelmed to use them, the "laggards".

This study estimates the causal effect of two scalable interventions on enterprise AI adoption: specialized short AI training sessions and email nudges that incentivize use and share relevant use cases. We pay particular attention to effects on persistent laggards, the subgroup most exposed to the labour-market risk that uneven enterprise AI adoption creates.

2026-05-18

2026-06-14

The primary outcome will be the intensity of GenAI use in the primary measurement window, measured by the number of OpenAI-tool messages sent by the employee.

The primary outcome will be the intensity of GenAI use in the primary measurement window, measured by the number of OpenAI-tool messages the employee sends. Taking into account the distribution of the data, we will decide the appropriate transformation, if needed, blind to treatment assignment. The second outcome will be the number of active days in the same window, defined as the number of distinct calendar days on which the employee sends at least one OpenAI-tool message.

Secondary outcomes are exploratory and cover adoption and intensity of use, breadth and depth of use, persistence of use, substitution across AI platforms, and self-reported attitudes and perceived value of AI.

The secondary outcomes are organized in five families, all computed on the primary measurement window unless otherwise noted.

Adoption and intensity: Adoption and intensity: any OpenAI usage, active days, active weeks, total messages, messages per active day, average conversation length, and winsorised messages.

Breadth and depth of use: the range of OpenAI tools used and whether multiple tools are combined; the use of advanced and agentic tools such as Codex, ChatGPT Agent, Deep Research, GPTs and Projects; connector use to workplace systems; and the task composition of use, for example education, coding, analysis and writing.

Persistence (weeks 5 to 8): Average active days per week, average active weeks, and sustained usage applying the persistent activation criteria to the same period.

Substitution: Gemini messages, total AI messages across platforms, and the OpenAI share of total AI messages.

Finally, we will register attitudes and perceived value of AIvia a short end-line with identical timing and channel across all arms, measuring self-reported attitudes towards AI, confidence, and perceived usefulness and time saving. Because response is voluntary and may differ across arms, this family is exploratory, and treatment effects on it are interpreted with that caveat.

The trial uses a 2x2 factorial randomised controlled design (Control, Nudge Only, Training Only, Nudge plus Training), which allows the effect of each of the two interventions, and the interaction between them, to be identified within a single experiment. Assignment is at the individual level, with equal allocation across the four arms. All comparisons are between randomly assigned arms, with no clustering since randomisation is at the individual level.

Not available

Random assignment at the individual level, with equal probability across the four arms, generated by computer.

Individual level randomization.

No

10,000 employees.

10,000 employees.

10,000 employees.