Evidence · Work
The month that became an afternoon
AI's most repeated promise is productivity. This page shows what happens when the promise is actually measured: in people working, with control groups and published results.
Those who knew least grew most
The largest field study ever run on generative AI at work followed 5,179 customer support agents at a software company. With an AI assistant suggesting replies, average productivity rose 14%. The number hides the main finding: for novice agents the gain was 34%, while for the most experienced it was close to zero.
The authors' reading is the one that matters: the assistant learned the pattern of the best agents and served it to those just starting out. AI worked as the veterans' experience, bottled. It leveled no one down; it shortened the path between day one and competence.
Figure 1
Productivity gain with an AI assistant, by agent experience
Novice agents
Overall average
Most experienced agents
Productivity measured in issues resolved per hour.
Source: Brynjolfsson, Li and Raymond, Quarterly Journal of Economics, 2025 (2023 NBER study). 5,179 agents.
Half the time, better work
In a controlled experiment with 95 developers, those who coded with an AI assistant finished the same task 56% faster: 71 minutes, against 161 for the group without AI.
In another experiment, with 453 college-educated professionals, professional writing tasks with AI took 40% less time and received 18% higher grades from evaluators who did not know who had used the tool. Less time and higher quality, measured together: the combination economics almost never delivers.
Figure 2
Time to finish the same programming task, in minutes
-56%
Without an assistant
With a code assistant
56% faster completion, with an equal or higher success rate.
Source: Peng et al., 2023. Controlled experiment with 95 developers.
Figure 3
-40%
time on professional writing tasks
+18%
quality, graded blind
453
professionals in the experiment
Source: Noy and Zhang, Science, 2023.
The pattern that repeats
Our About page says we want people who deliver in an afternoon what used to take a month. The sentence is not rhetoric: it is what these three studies measured, each in its own way, in support, code and writing.
And notice what none of them found: replacement. They found novices competent sooner, professionals delivering more and higher-quality work. The difference between using AI to replace people and using it to multiply them is not technical, it is a choice. Ours is made.
