Scratch that. Here it is : Large Language Models, Small Labor Market Effects by Anders Humlum, Emilie Vestergaard :: SSRN
Summary, courtesy of Chat GPT 
AI has not significantly improved productivity — at least not yet:
A study of 25,000 Danish employees (across 7,000 organizations) found no meaningful impact on wages or working hours.
Adoption has been surprisingly fast, but the benefits aren’t visible in monetary terms.
AI is creating more new tasks than it eliminates:
For example, teachers often face more work due to students secretly using AI — requiring extra oversight and judgment.
Prompt writing and validating AI-generated outputs also consume time.
Productivity gains remain modest:
Average time savings were only 2.8%.
While lab studies have shown 15–50% gains, these results have not translated into real-world environments.
In some cases, AI may reduce work quality:
According to Boston Consulting Group, some employees begin relying too heavily on AI and stop thinking critically themselves.
Large companies are tapping the brakes:
Johnson & Johnson tested AI in 900 use cases — only 10–15% generated around 80% of the benefits.
Salesforce and The Information report that most AI investments haven’t yielded significant financial returns.
FOMO drives adoption despite weak results:
In a recent IBM survey of 2,000 executives, only 25% said their AI initiatives had met ROI expectations.
Still, optimism remains high for long-term benefits.
Conclusion:
AI has not yet lived up to the high expectations placed on it. Its impact has been slower, narrower, and harder to quantify than public discourse suggests. Rather than eliminating work, it often creates new kinds of tasks — and in the wrong context, it can even degrade performance. Companies fear missing out on the AI wave, but few have figured out how to use it profitably.