Tenure and output

GAICIL Insights

Economists have long read learning on the job from paychecks, assuming pay tracks productivity.¹ Tenure and output, by Kathryn Shaw and Edward P. Lazear, checks that assumption with something rare: data on both wages and individual output. The setting is an autoglass company, where most production happens at the individual level and output is clearly measured.

The paycheck turns out to be a terrible thermometer. The abstract reports a very steep learning curve in the first eight months: output is 53 percent higher after eight months than at the start. Pay lags far behind. Pay profiles are much flatter than output profiles in the first year and a half. Under the hourly wage regime there is no pay gain with tenure in the first 19 months, and under the piece rate regime the gain is modest relative to output. Wage profiles significantly underestimate human capital investment compared with output-tenure profiles. The authors call this the most important point: wage profiles are poor proxies for output profiles.

Selection matters too. Workers who stay longer have higher output and faster early learning. For older workers, after the first two years of tenure, productivity declines with additional tenure. The authors favor declining effort, perhaps from boredom, as the likeliest explanation, since skills hardly vanish that fast.

Now ask what artificial intelligence does to that steep early curve. AI could push it in either direction. It could steepen it, giving new workers guidance that speeds skill acquisition. It could flatten it, doing the tasks workers would otherwise master through practice. Since learning is so rapid and large in the first months, AI's effect on that window could be enormous. The first eight months of a job may be where the AI productivity story is quietly written.

The pay-output gap creates a measurement blind spot. If wages understate early productivity growth, wage-based returns to experience miss much of the learning. As AI changes how quickly workers become productive, wage data may miss that change too. Firms and researchers need output data to see AI's effect on learning, just as this paper needed output data to see tenure effects. Studies that judge AI by earnings alone may be looking at the wrong gauge.

The late-tenure decline raises a design question. If boredom erodes effort among long-tenured workers, AI could help by automating repetitive tasks and leaving more varied work. Or it could worsen monotony by reducing jobs to supervising automated processes. The effect on effort, and hence productivity, depends on how jobs are rebuilt around AI. A tool that removes the interesting parts of a job may carry hidden costs.

Policy implications follow. On-the-job learning is a major productivity source that wage statistics understate. Encouraging firms to hire and train new workers, and supporting the early months of employment, may yield larger gains than wage data suggest. If AI tools accelerate early learning, supporting their use in training contexts could raise productivity substantially.

Compensation design is the final piece. The pay-productivity link differed between hourly and piece rate regimes, with piece rates tracking productivity more closely. As AI makes individual output easier to measure, firms may move toward pay that rewards output more directly. Such systems could align pay with the rapid early learning documented here, rewarding workers as they improve and giving them a stake in AI-assisted gains.

Seen whole, the paper's message is plain: learning on the job is steep, fast and poorly captured by wages. Measuring it requires output data. As AI reshapes tasks and learning speed, researchers and policymakers will need similar data to understand its effects. The autoglass workers show how much productivity grows in a job's early months. Making sure AI supports that growth, and that workers share in its rewards, is a central challenge for the future of work.

A final provocation: the real return to AI-assisted training may be invisible in payroll records for a year or more. Anyone evaluating such programs by wages alone should expect to underestimate them.