👋 Hey, it’s Stephane. I help engineers become great engineering managers - whether you want to become one or are already leading a team.
Paid subscribers get 50 Notion Templates, The EM’s Field Guide, and access to the complete archive.
Your team metrics might be looking great. Compared to previous periods you see more code being shipped, faster time to first PR review, faster PR cycle time, and less incidents. Every metric you were told to watch is moving in the right direction.
Now, have a look at your roadmap… The features your team was supposed to deliver this quarter are landing on roughly the same schedule they always did.
What’s going on? Surely all the metrics you see should be moving the needle on your deliver somehow!
The individual productivity gains
Mert Demirer, Leon Musolff, and Liyuan Yang looked at GitHub activity against AI-usage telemetry for more than 500,000 developers, tracking three generations of tools:
autocomplete
interactive coding agents
autonomous coding agents
Their NBER paper found that developers using all three generations of AI coding tools produced 240% more commits.
But the increase was much smaller further along the development process: they worked on 80% more projects and produced only 30% more releases.
Researchers call this a weak-link problem. AI speeds up one stage that was not the constraint, and leaves every human-dependent stage (reviews, integration, release) taking just as long as they did.
Microsoft found the same thing.. Sudhakar Sadasivuni, a principal group engineering manager at Microsoft, said: “We quickly identified that improving the individual productivity of a developer was not resulting in a boost to team productivity. That was our hard lesson.”
A separate analysis of 623 million code changes, from GitClear, points at the same gap. As AI-generated code volume rose, cross-file reuse and refactoring fell, and long-term maintenance work dropped by nearly three-quarters compared to 2022 levels.
This is not really a surprise
I’ve made a version of this argument before about performance reviews: the team is the only honest unit you can measure, because individual numbers can move in the opposite direction from what the team delivers. Obviously metrics like counting PRs, or commits fall into the same trap.
SDLC still has to move through several stages: planning, coding, review, integration and release. Making one stage faster doesn’t automatically speed up the rest.
If your review stage was already the slowest stage, AI actually may make the problem worse. More code is now arriving, but the team’s capacity to review it hasn’t changed.
The good news is that you can find and fix these bottlenecks.
Since writing code is no longer the slowest part for most, what is? Here’s how to find out, and what to do about it:
👇👇👇 (access with a paid subscription)