Claims that AI tools make people dramatically faster are everywhere, but the best evidence is more specific, and more useful, than the headlines. Two studies from the National Bureau of Economic Research (NBER), a U.S. research organization, point in what looks like opposite directions, and reading them together shows where AI helps a team and where it does not.
Two studies, two very different questions
The first paper, “Generative AI at Work” by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, looks inside one company. It studies a conversational assistant built on generative AI (software that produces text in response to a prompt) that was rolled out in stages to 5,179 customer support agents. According to the NBER, it was issued in April 2023, revised in November 2023, and later published in The Quarterly Journal of Economics in 2025.
The second paper, “Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI” by Anders Humlum and Emilie Vestergaard, looks at a whole labor market. It links worker surveys on chatbot adoption to Danish administrative records. According to the NBER, it was issued in May 2025 and revised in March 2026, and it was earlier circulated under the title “Large Language Models, Small Labor Market Effects.”
One asks, “Does this tool help the people using it do a specific job faster?” The other asks, “Has it changed what workers earn or how many hours they record across the economy?” Those are different questions, and they can have different answers.
What happened inside the call center?
The staged rollout is the most useful feature of the first study. Because some agents received the tool before others, the researchers could compare similar people with and without access at the same time, rather than guessing from before-and-after snapshots.
According to the NBER abstract, access to the tool raised productivity, measured as issues resolved per hour, by 14% on average. The authors also report better customer sentiment, higher employee retention and possible learning by workers.
The gains were not spread evenly. Novice and low-skilled workers improved by 34%, while experienced and highly skilled workers saw minimal impact. The authors suggest the tool spreads the practices of more capable workers and helps newer staff gain experience faster.
A worked example of the arithmetic
Percent gains in output per hour are easy to misread, so an illustrative calculation helps. Suppose a support agent resolves 5 issues per hour before the tool.
| Scenario | Gain | Issues per hour | Time per issue |
|---|---|---|---|
| Before the tool | none | 5.0 | 12.0 minutes |
| Average gain | 14% | 5.7 | about 10.5 minutes |
| Novice-level gain | 34% | 6.7 | about 9.0 minutes |
The figures follow from simple multiplication: 5 × 1.14 = 5.7 and 5 × 1.34 = 6.7 issues per hour, and 60 divided by each rate gives the minutes per issue. Note that a 14% increase in output per hour is not a 14% cut in time. Because 1 ÷ 1.14 is about 0.877, each issue takes roughly 12% less time, not 14%.
Why would a country-wide study find so little?
The second study took a much wider view. According to the NBER, the authors estimate null effects, meaning no measurable change, on earnings and recorded hours at both the worker and workplace levels. They rule out effects larger than 2% two years after ChatGPT’s launch. The method is difference-in-differences, which compares how outcomes changed for adopters against how they changed for similar non-adopters over the same period.
That does not mean workers found the tools useless. The abstract notes that workers report productivity benefits. It also says that adopters move into higher-paying occupations, but that these shifts are still too few to move average earnings. It adds that employers absorb the technology by reorganizing tasks, with new work appearing in content generation, AI oversight and AI integration.
In plain terms, a time saving on one task does not automatically show up as a pay change or fewer hours worked. The freed time may be spent on other tasks, on checking the tool’s output, or on new work the tool itself creates.
How can both results be true?
Several differences explain the gap.
- Scope. One study measures a single task in a single firm with a deliberately structured rollout. The other measures pay and hours across many workplaces and occupations.
- Outcome. Issues resolved per hour is a direct productivity measure. Earnings and recorded hours are far removed from it, and many things other than output change them.
- Work type. Customer support follows a repeatable pattern, with many similar conversations, which is what a text-generating assistant handles well. Much knowledge work is less uniform.
- Time. Task reorganization and pay adjustments tend to lag the first productivity gains.
A short-range, task-level gain and a flat economy-wide average can coexist for a long time. Neither study, on its own, tells a manager what will happen on their team.
What this means for a team deciding whether to adopt AI
Three practical lessons follow from reading the papers side by side, without going beyond what they report.
- Measure a task, not a feeling. The call center study worked because it tracked a concrete rate. Teams can do the same on a small scale by picking one repeatable task and tracking output per hour before and after, along the lines of what output per hour tells your team.
- Expect uneven benefits. The 34% versus minimal split suggests tools may help people with less experience most, while experts gain little. That affects who should get training first and how to judge results fairly.
- Plan for the new work. If AI oversight and integration become tasks of their own, time saved in one place is partly spent in another. Count that cost when estimating the net gain.
Anyone new to the technology can start with what generative AI is, and the governance side is covered in using AI responsibly in your business.
Where the evidence stops
Both studies have limits. The first covers one company and one type of job, so its numbers should not be applied to, say, software design or legal drafting. The second is based on Danish data and measures outcomes up to two years after the launch of a particular chatbot, so it cannot speak to later tools or longer horizons. The authors of the second paper rule out large effects on pay and hours but do not claim that every worker is unaffected.
Tools also change quickly. Any specific percentage, including the 14% and 34% above, describes one setting at one time. Treat such numbers as a reason to run a careful local trial, not as a forecast.
A simple way to test AI on your own work
A lightweight trial borrows the logic of the call center study without needing a research budget.
- Pick one repeatable task with a clear finish line, such as drafting a standard customer reply or summarizing a weekly report.
- Record a baseline for a couple of weeks: items finished per hour and how often work needs correction.
- Introduce the tool to part of the group first, so there is a comparison rather than just a before-and-after.
- Track quality alongside speed. The call center study also looked at customer sentiment, a reminder that faster is only better if the result holds up.
- Review after a fixed period and keep the tool only where the numbers and the people doing the work agree it helps. Pairing the trial with a simple system for managing tasks keeps the comparison honest, because everyone is counting the same things.
Reading AI productivity claims with a steady eye
The two papers offer a balanced lesson: measured, task-level gains can be real and sizable for some workers, while broad economic averages move slowly or not at all. When a claim about AI and productivity appears, ask what exactly was measured, in which job, for whom, and over what period. A team that answers those questions for its own work will learn more than any headline can tell it.
This article provides general information for readers interested in work and productivity and is not business, financial or legal advice.
Sources: National Bureau of Economic Research
