Member-only story
Part 8: AI in Maintenance and Developer Experience, and How to Know if AI Is Actually Helping
The final part of “AI Across the SDLC”: tech debt, developer friction, and the measurement mistakes that make AI look more helpful than it is.
In July 2025, the research group METR ran a careful experiment. Experienced open-source developers worked on real tasks in codebases they knew well, sometimes with AI tools and sometimes without. With AI, they took 19% longer. The surprising part: they believed AI had made them 20% faster.
That’s a 39-point gap between how productive people felt and how productive they actually were.
The story didn’t end there. METR expanded the study in a follow-up reported in February 2026, with 57 developers, 143 repositories and more than 800 tasks, and the results moved. METR now says developers are likely getting more speed from AI tools in early 2026 than their 2025 numbers suggested. So the 19% figure shouldn’t be quoted as a current fact.
But the bigger lesson still stands: how fast AI makes us feel isn’t evidence of how much it’s helping. That’s the right place to end this series, because the final stage of the SDLC, maintaining software and keeping developers productive, is where AI’s real value either adds up or quietly disappears.
