Before asking whether AI pays, decide what counts as a payoff

Imagine a team finishing its first draft sooner, then using the remaining time for another assignment. That could be a useful change. Establishing what happened to costs, output or pay would require further measurement. A single word such as payoff can conceal several questions.

TechCrunch covered Apollo’s assessment of AI returns on October 25. In his October 24 note, Apollo chief economist Torsten Slok interprets revenue per worker and profit margins as evidence that financial benefits remain concentrated among a leading technology group. That is his reading of aggregate indicators, rather than a direct estimate for each adopting business.

Give the outcome a precise name

The U.S. Bureau of Labor Statistics defines labor productivity as real output divided by hours worked. Its broader total-factor measure relates output to a combination of inputs. Neither definition is simply a count of how many people have opened a chatbot or how quickly one draft appeared.

The OECD adds a useful caution: output per hour reflects how labor works with other inputs, including capital, materials and organizational change. It only partly reflects workers’ own capabilities or effort. An observed productivity change therefore calls for an explanation of the inputs and circumstances behind it.

These definitions help separate adoption, task performance and business results. More use of a tool establishes activity. To evaluate a claimed gain, the next questions concern the output produced, the resources required and the period over which they were compared.

Keep the study’s window in view

In an October 6 research account published by Brookings, Anders Humlum and Emilie Vestergaard describe surveys of about 25,000 Danish workers in 11 AI-exposed occupations, linked to administrative records through December 2024. They compared chatbot adopters with similar non-adopters and reported no detectable average effects on earnings or recorded hours in that early window.

The same authors describe changes in tasks and work organization. Their findings concern particular labor-market outcomes in Denmark during the observed period; they do not measure every company’s profits or establish the current payoff from U.S. deployments. The distinction between a reported change in work and a measured change in pay is central to interpreting the evidence.

Make a local claim possible to check

For a hypothetical team evaluating a drafting assistant, a useful measurement plan could specify the kind of document, the required quality and the work counted. Time spent checking, correcting and integrating a draft would belong alongside generation time. The comparison could then ask what changed across the complete assignment, with the same standard for an acceptable result.

A second question would concern how any recovered time was used. Did the team finish more accepted assignments, spend longer reviewing them or change its working hours? These are proposed evaluation questions, rather than findings about a particular product, workplace or study.

Broad economic assessments and detailed workplace research can inform that exercise while answering different questions themselves. A clear AI payoff claim identifies an outcome, a population and a comparison period. That makes both an encouraging result and an inconclusive one easier to understand—and easier to revisit when better evidence arrives.