A business that asks an employee to help train a system intended to replace their job faces an incentive problem that could affect the speed and cost of automation, according to a new economic model.
In an October working paper published by the National Bureau of Economic Research, Joshua S. Gans and Peter Landry examine a situation in which a worker’s cooperation is needed to train an AI system, but successful training ends the worker’s job.
Their analysis is theoretical. It does not measure how often employees resist AI adoption, establish that most AI systems replace jobs, or predict the outcome at a particular company. It does show why the terms offered to workers can matter when an employer depends on their expertise.
Why an employee’s knowledge can matter
Automating a task often requires someone to explain how work is done, identify exceptions and assess whether a system’s outputs are usable. An experienced employee may know which cases need special handling or why a seemingly reasonable answer would cause a problem.
Imagine a business developing a tool for routine customer inquiries. A member of staff helps label examples and checks the proposed responses. If the employee expects the project to remove their position, the request to assist can create an obvious conflict between the company’s goals and the worker’s interests.
This is a hypothetical example, not an observation from the paper. Many AI projects aim to support employees or change part of a job rather than eliminate the position. The model applies to the more specific case where training succeeds and the worker’s job ends.
What the model examines
Gans and Landry model a worker who can cooperate with training, withhold effort or obstruct training while continuing productive work. In their framework, firing the worker also means the firm loses the help required to complete the training.
That makes dismissal a weaker deterrent to withholding cooperation. The authors find that the employer may have to give the worker more value to secure cooperation, and that a firm relying only on wages may accept a slower training process.
Economists use the term worker rent for the value an employee receives above their available alternative. It is not necessarily a cash bonus. It can reflect how valuable the job is to the worker compared with their next best option.
These are results within a model built on particular assumptions about the worker, employer and training process. They should not be read as a description of how most employees behave.
Why a completion payment could change the calculation
The authors also consider a credible payment made when training is completed. In their model, a completion bonus equal to the worker’s rent without AI removes both the worker’s additional gain and the slowdown.
The word credible is essential. A promised payment has to be structured so that the employee can reasonably expect to receive it under the agreed conditions. The paper does not establish a universally appropriate bonus or show that a bonus alone will resolve workplace concerns.
Real employers may also consider redeployment, training for another role, transition support or a negotiated timeline. Those are possible approaches, not interventions tested in the paper. Their suitability and legal implications depend on the company and jurisdiction.
Automation costs extend beyond the software
For managers, the paper suggests separating the price of a tool from the cost of making it useful. An implementation budget may need to account for employee time, review, documentation, transition arrangements and the consequences of changes to staffing.
Output quality is another consideration. A system may learn routine cases while struggling with uncommon requests. An employer needs a way to test those cases and decide who remains responsible for errors.
Our coverage of how companies organize employee training examines a related question: what workers learn and how a business uses that knowledge depend on the way work is arranged. Here the knowledge transfer serves a different purpose, and the employee’s future is part of the arrangement.
What the research cannot tell employers
The paper does not establish how a particular workforce will respond, whether a proposed automation project will work, or whether a company will actually reduce employment. It does not compare real-world bonus schemes.
Its value is to make an incentive visible before an implementation plan is treated as a simple technical project. If a business needs people to teach a system that may change their roles, it must consider what those people stand to gain or lose, and what commitments the company can realistically keep.
