Large-Scale Study: AI Saves Just 3% of Work Hours, With Minimal Earnings Impact
Summary
- • Danish study of 25,000 workers finds AI saves only 2.8% of work hours on average
- • Just 3–7% of AI-driven time savings translate to higher earnings or company profits
- • Lab studies show 15–55% task speedups that largely vanish in real-world conditions
- • Gains are genuine for writing and support tasks but evaporate across full workdays
Details
2.8% work hours saved
Study of ~25,000 workers across 7,000 Danish workplaces found AI adoption saves approximately 2.8% of work hours — roughly one hour per week — per actual payroll records.
3–7% earnings conversion rate
Only 3 to 7% of productivity gains from AI adoption reached workers' actual earnings or recorded hours. AI had no significant impact on earnings or hours across all occupations studied.
40% faster writing, 18% quality boost
Randomized experiment with 453 professionals giving ChatGPT access for mid-level writing tasks (press releases, reports, emails) cut time by 40% and raised graded output quality by 18%.
14% support productivity gain
Field study of 5,179 customer-support agents found AI assistant access lifted resolved-issues-per-hour by 14% on average, and approximately 34% for the newest, least experienced agents.
Harvard/BCG: 19% penalty off-frontier
Among 758 consultants using GPT-4, those using AI were 19 percentage points less likely to reach correct answers on tasks deliberately outside AI's competency range vs. those working without AI.
Lab vs. real-world gap
Lab studies measure a single controlled task (showing 15–55% gains). Real-world studies measure impact across full months and payrolls. The gap exists because AI doesn't touch most of a worker's job.
ROI requires deliberate capture
Saved time doesn't convert to revenue automatically. Workers and businesses must consciously redirect freed hours into billable work, more clients, or cost reductions — or the gain evaporates.
Evidence from controlled studies and large-scale payroll data on where AI productivity gains are real, where they fail, and why they rarely reach the bottom line.
What This Means
This large-scale Danish study is among the most rigorous real-world assessments of AI's economic impact to date, linking AI adoption directly to payroll records rather than controlled lab settings. The finding that only 3–7% of AI productivity gains reach actual earnings directly challenges the widespread assumption that AI tools automatically improve business performance. For organizations investing heavily in AI productivity tools, this research signals that unlocking genuine financial returns requires intentional workflow redesign — deliberately converting freed time into billable output — rather than passive deployment and hope.