Engineering video
1:16 · 20 August 2025
The speaker explains: studies of millions of commits in real-life enterprise software show that AI increases productivity by 0 to 30 percent, but much of that gain is offset by the need to constantly rework, retouch, and clean up generated code in subsequent commits.
Hear it at 0:00 →we finally got some real tangible empirical analysis of what AI and how much AI can actually help us as developers. So in the study they analyzed over 100,000 software engineers from 600 companies and millions of commits. And they tried to measure productivity not based on lines of code or commits, but actually on functionality and code quality. And this was focused on real-life enterprise software, not hobby project or greenfield apps.
What they saw it's a real productivity increase of 0 to 30% depending on the task and the complexity of the task, the size of the code base and the language. But what they also observed is that whenever you're using AI, you spend a lot of time reworking code. And the reason for that is because yes, you can generate a lot more code, but you'll have to edit that code, you'll have to do what they called rework.
Basically in the following commits or a couple of spins later, a lot of that work gets redone or cleaned up. Uh this is not refactoring, is not going somewhere and rebuilding that, we're making it better. It's actually re retouching work that you've just done. And so if you compute that difference, you realize you're only 10 to 20% faster.
