Measured benefits of AI come in task-sized pieces. In a study of 5,179 customer support agents, an AI assistant raised the number of issues resolved per hour by 14 percent on average, and in a preregistered experiment with 453 professionals, ChatGPT cut the time spent on writing tasks by 40 percent. A Danish study that linked chatbot adoption to government payroll records found no detectable change in earnings or recorded hours, so task-level gains had not reached paychecks as of its March 2026 revision.
Key takeaways
- Controlled studies report real gains in customer support (14 percent more issues resolved per hour) and in short professional writing tasks (40 percent less time).
- The largest customer support gains, 34 percent, went to novice and low-skilled agents, while experienced agents saw minimal change.
- A machine learning weather model, GraphCast, beat the most accurate operational deterministic system on 90 percent of 1,380 verification targets in a peer-reviewed comparison.
- Danish administrative data rule out earnings or hours effects larger than 2 percent two years after ChatGPT launched.
- Several results from 2023 and 2025 have since been revised, retitled, or labeled out of date by their own authors, so each figure needs its date attached.
Customer support: 14 percent more issues resolved per hour
Economists Erik Brynjolfsson, Danielle Li, and Lindsey Raymond studied a generative AI conversational assistant that a company introduced to its customer support agents in stages. Access arrived at different times for different people, so the researchers could compare agents with and without the tool over the same period. Across 5,179 agents, productivity measured as issues resolved per hour rose by 14 percent on average. The National Bureau of Economic Research working paper, later published in The Quarterly Journal of Economics in 2025, reports that novice and low-skilled workers gained about 34 percent, while experienced and highly skilled workers saw minimal impact. The authors also report better customer sentiment and higher employee retention.
The scope is narrow. One company's support operation and one tool were studied, so the result describes that setting and nothing wider.
Writing tasks: faster, and rated higher
Shakked Noy and Whitney Zhang ran a preregistered online experiment in which 453 college-educated professionals received incentivized, occupation-specific writing tasks. Half were randomly given access to ChatGPT. Average time taken fell by 40 percent and output quality rose by 18 percent. Differences between workers narrowed. Two weeks later, people who had used the tool were twice as likely to report using it in their real jobs, and two months later they were 1.6 times as likely. The paper appeared in Science in July 2023.
The tasks were short and mid-level, and an experiment has no deadlines, colleagues, or accountability. The result says a first draft can go faster. A whole job becoming 40 percent faster would be a different claim, and the paper does not make it.
Weather forecasting: a benefit measured outside the workplace
GraphCast, a machine learning model described in Science in November 2023, is trained directly on reanalysis data and predicts hundreds of weather variables for the next 10 days at 0.25-degree global resolution in under one minute. In the authors' evaluation it significantly outperformed the most accurate operational deterministic forecasting system on 90 percent of 1,380 verification targets, and it improved tracking of tropical cyclones, atmospheric rivers, and extreme temperatures.
The comparison is against one class of forecast, the single best-estimate run, on a fixed set of test targets. A model that wins on most targets can still lose on a particular storm, which is why forecasters keep comparing methods over time.
Danish payroll records: where pay did not move
Anders Humlum and Emilie Vestergaard linked large surveys on AI chatbot use to administrative labor market records in Denmark. Using a difference-in-differences design, they found precise null effects on earnings and recorded hours at both the worker and workplace level, ruling out effects larger than 2 percent two years after ChatGPT's launch. The NBER working paper notes that most employers in exposed occupations have adopted chatbot initiatives and that workers report productivity benefits. Employers reorganized tasks and created new work in content generation and AI oversight, and some adopters moved into higher-paying occupations, but too few did so to shift average earnings.
The paper was issued in May 2025 as "Large Language Models, Small Labor Market Effects," and the March 2026 revision carries the title "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI." The NBER page lists no journal publication.
The finding sits alongside the experiments without contradicting them. A worker who drafts faster may spend the saved time on other tasks, and an employer may leave pay unchanged, so a task-level gain can stay invisible in wages.
What has changed since these studies appeared
Weather is the clearest case of a result moving into routine use. ECMWF, the European intergovernmental forecasting center, put its machine learning model AIFS Single into operation on February 25, 2025, at 28 km grid spacing, and says it complements its physics-based system rather than replacing it. An ensemble version, AIFS ENS, followed in 2025 with 50 perturbed members plus one control. ECMWF reports gains of up to 25 percent for upper-air variables against its physics-based ensemble, while the older system stays more skillful for 10-meter wind speed.
Software work gave a cautionary example. In a 2025 randomized trial by the research group METR, 16 experienced open-source developers took 19 percent longer on 246 real issues when AI tools were allowed, after predicting a 24 percent speedup and later still believing they had been about 20 percent faster. METR now marks those results as out of date and published newer data in February 2026 on late-2025 tools. The episode shows how fast a measurement ages when the tools change every few months.
Side-by-side summary of the studies
| Study | Setting | Reported result | Main limit |
|---|---|---|---|
| Brynjolfsson, Li, and Raymond | 5,179 support agents, staged rollout | 14 percent more issues resolved per hour; 34 percent for novices | One company and one tool |
| Noy and Zhang | 453 professionals, online experiment | 40 percent less time, 18 percent higher quality | Short tasks, not full jobs |
| GraphCast team | Global weather forecasts, 10-day range | Better on 90 percent of 1,380 verification targets | Compared with one forecast type |
| Humlum and Vestergaard | Danish survey and payroll records | No earnings or hours effect above 2 percent | Covers two years after launch |
| METR developer trial | 16 developers, 246 issues | 19 percent longer with AI allowed | Labeled out of date by METR |
Questions to ask about any claimed benefit of AI
- Who was measured, how many, and for how long?
- Was there a comparison group, or only a before-and-after look?
- Is the outcome time, quality, cost, or earnings, and who measured it?
- Did the workers or organizations that benefited resemble the ones the claim is about?
- Which version of the tool was used, and has the source been revised or retired since?
FAQ
Does AI make every worker more productive?
No. In the customer support study, experienced agents gained little while novices gained the most. Effects depend on the task, the person, and how the tool is used.
Why would pay not change if productivity rises?
Pay depends on more than output per hour. The Danish study found that employers reorganized tasks and that too few workers moved into higher-paying occupations, so average earnings stayed flat within the two years observed.
Are the 2023 results still current?
They are accurate measurements of the tools available then, and none of the main papers has been withdrawn. The Danish paper was revised in March 2026 and METR retired its own 2025 figures, so newer tools may produce different results. Each number belongs with its date.
Is a benefit in one field proof of benefit in another?
No. A forecasting model that beats a physics-based system on test targets says little about whether a writing assistant helps a nurse or a teacher. Each setting needs its own evidence.
Related reading
- Advantages and impacts of automation and AI
- Weighing the pros and cons of AI in society
- What artificial intelligence is and where it is applied
- Well-known AI systems around the world
More in AI Basics.
Last verified: October 2026
Sources
- Generative AI at Work, Brynjolfsson E, Li D, Raymond L, NBER Working Paper 31161, 2023; The Quarterly Journal of Economics 140(2), 2025
- Experimental evidence on the productivity effects of generative artificial intelligence, Noy S, Zhang W, Science, July 2023
- Learning skillful medium-range global weather forecasting, Lam R et al., Science, November 2023
- Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI (earlier title: Large Language Models, Small Labor Market Effects), Humlum A, Vestergaard E, NBER Working Paper 33777, May 2025, revised March 2026
- ECMWF's AI forecasts become operational, ECMWF, 2025; AIFS ENS becomes operational, ECMWF Newsletter 185, 2025
- Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR, July 2025 (marked out of date, February 2026)