The evidence base · 4 sources · 2 min read · Read with care
AI in the machine, 19 August 2026 · Sources read 25 June 2025 to 28 August 2026
2.8%
of users' working hours saved by AI, in a study of roughly 25,000 workers across 7,000 workplaces
The number that should temper everything
The largest rigorous study of AI and time saved covers roughly 25,000 workers per round across about 7,000 workplaces and 11 occupations, including marketing professionals and office clerks.
Average time saving among the people using the tools: 2.8% of their working hours. Precise null effects on earnings and recorded hours.
Of the new tasks AI created, 59% were implementing and overseeing the AI itself.
Alongside it: Deloitte found 68% of organisations had moved 30% or fewer of their generative AI experiments into production, in the third quarterly edition of its enterprise survey. Gartner reckons roughly 130 of the thousands of vendors calling themselves agentic AI are genuine.
None of that means AI does nothing. It means the gap between what is announced and what is running is very large, and there is no published evidence that this sector beats the base rate, in either direction.
Why this is labelled read with care
- These studies are economy-wide, not sector-specific. There is no published measurement of AI and productivity inside affiliate, reward or creator platform operations - in either direction. The honest position is that the evidence is thin both ways, and the sceptical case is stronger mainly because it is better sampled.
- I have deliberately not used the widely quoted "95% of pilots fail" figure. It rests on 52 interviews, is not peer reviewed, and measures something narrower than it is usually quoted as measuring.
Sources
Humlum and Vestergaard, Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI, NBER working paper 33777, revised March 2026Read at source 28 August 2026
The precise null effects on earnings and recorded hours. The 2.8% and the task split come from the earlier version, cited separately below.
Humlum and Vestergaard, Large Language Models, Small Labor Market Effects, Becker Friedman Institute working paper 2025-56Read at source 28 August 2026
The 2.8%, the 25,000 workers across 7,000 workplaces, and the split of new tasks, all in 1 document.
Deloitte, The State of Generative AI in the Enterprise: Now decides Next, third edition, August 2024Read at source 19 August 2026
The 68% who had moved 30% or fewer experiments into production.
Declared interest: Deloitte sells AI implementation services.GartnerRead at source 25 June 2025
The estimate of roughly 130 genuine agentic AI vendors.
Declared interest: Gartner sells vendor evaluation.
Every source carries the date it was read. Where there is no link the document is held on file and cited by publisher and title.
My read
The useful reading is not that AI does nothing. It is that the distance between what gets announced and what is actually running is very wide everywhere, so planning on the base rate is reasonable until you have measured your own.
What I'd do with this
Before you buy an AI tool, write down what it is supposed to remove and how you would measure that in 90 days. If you cannot write the measurement, you will not be able to tell whether it worked, and neither will the vendor.
Can your operations carry the growth you're planning?
Follow a transaction from the moment it happens to the moment everyone has been paid. Then find the place it stopped.
Get in touchFind where the money stands still