In late 2024, Danish researchers working with Statistics Denmark surveyed about 25,000 workers across 7,000 workplaces in 11 chatbot-exposed occupations (NBER w33777, March 2026), among them accountants, customer support specialists, software developers and teachers. Many teachers reported a new chore: detecting AI-generated homework. They then linked each answer to official earnings and hours records.
Who is responsible
Employers decided how this rollout went. In the study by Anders Humlum and Emilie Vestergaard (NBER Working Paper 33777, revised March 2026), about 43% of employers in these occupations explicitly encouraged chatbot use and about 6% banned it. Where the employer neither encouraged use, supplied a company chatbot nor trained staff, about 41% of workers had still used one at work, the same revision reports. Where all three were in place, the revision finds 93% had, 28% used one daily, and 19% said it saved them more than an hour a day.
So the employer's choices shaped how much time a worker saved. That matters for the next question.
Who benefits and who pays
The researchers linked the survey to monthly payroll records through December 2024, two years after ChatGPT launched (NBER w33777, March 2026). They compared adopters with similar non-adopters before and after chatbots arrived, a method called difference-in-differences. In the March 2026 revision of NBER Working Paper 33777, their estimates for earnings, hours and hourly wages all centre on zero, and they rule out average effects larger than 2%.
The time savings are real. That is true. They have not reached the payslip. In the same revision, the null held for daily users, for workers reporting over an hour saved a day, and for workplaces with the full package of encouragement, tools and training. It held in each of the 11 occupations (NBER w33777, March 2026).
Then where did the time go? In the same revision, 85% of chatbot users said they moved saved time to other tasks, and fewer than 10% took more breaks or leisure. Among users at workplaces with active chatbot initiatives, roughly 18% took on entirely new tasks (NBER w33777, March 2026), such as reviewing AI output, handling compliance and wiring the tools into workflows.
Efficiency is not the same as welfare. On this evidence, the gain was absorbed as more work inside the firm. It did not show up as more pay or fewer hours for the worker.
What the evidence says, and its limits
NBER w33777 (March 2026 revision) covers one country, eleven occupations and the first two years. It is a working paper, not yet peer-reviewed. The authors found no effect even in occupations with more flexible wage-setting. They also found that adopters were more likely to switch into higher-paying roles where chatbots matter more. In the March 2026 revision, those switchers' earnings grew 12 percentage points faster than other employed workers, but there are still too few of them to move the average.
The firm-level evidence differs. In a staggered rollout to 5,179 customer support agents, an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers (Brynjolfsson, Li and Raymond, Quarterly Journal of Economics, 2025). Two years on, Denmark's national payroll data still show no change in pay or hours (NBER w33777, March 2026).
Here is what would change my mind: payroll records from 2025 onward showing adopters' earnings or hours pulling away, or bargaining data showing saved time being priced into pay.
The fix: decide where the hour goes
Nobody decides where a saved hour goes, so by default it turns into more work. Leaders can make it a decision:
- Measure it. Track time saved per role and task, before and after rollout, not just usage.
- Name the destination. For each role, decide what share of saved time goes to pay, to training, to lighter workload, or to new output.
- Price the new tasks. Review of AI output, compliance and integration are real work. Put them in job descriptions and grading.
- Count the non-users. Check who picked up AI-related work without using the tools.
- Report back. Share the numbers with staff or their representatives each quarter.
What This Means for You
Workers: Ask three questions. How much time does this tool save in my role, and who measured it? Where is that time going: more output, new tasks, training or pay? Is the AI review work I now do in my job description?
Managers: Log the new tasks your team takes on. Unlogged work can't be paid or planned for.
Leaders: Employer initiatives roughly doubled take-up in NBER w33777 (March 2026 revision). Decide, in writing, who shares the gain.
Technology is only progress if it moves people forward.
References
- Humlum, A. and Vestergaard, E. "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI" (previously "Large Language Models, Small Labor Market Effects"). NBER Working Paper 33777, May 2025, revised March 2026. https://www.nber.org/papers/w33777
- Brynjolfsson, E., Li, D. and Raymond, L. "Generative AI at Work." NBER Working Paper 31161 (2023); Quarterly Journal of Economics 140(2), 2025. https://www.nber.org/papers/w31161