Just finished a six-week tracking exercise on my team's Copilot usage. We measured code churn—the percentage of lines added or deleted in a commit versus the total lines changed. The average increased from 35% to 55% since the org-wide rollout. That's a 20-point jump.
For those not familiar, high churn often means more rewriting, more trial and error, more back-and-forth. Copilot is fantastic at generating blocks of code quickly, but that speed comes with a cost. I'm seeing developers accept a suggestion, run it, then immediately iterate or delete large portions of it because the logic wasn't quite right or the abstraction was off. It's creating a fast-type, backspace, fast-type rhythm.
I'm torn. Velocity on boilerplate and common patterns is definitely up. But I'm questioning the net effect on thoughtful architecture. When you're iterating that fast, you're less likely to pause and consider if you're building the right abstraction or if that module's responsibilities are bleeding. This feels particularly acute in infrastructure-as-code (Terraform) and B2B service boundaries, where a wrong pattern can be costly to fix later.
Has anyone else done similar metrics? I'm looking for a signal on whether this is just an adjustment period—developers learning to prompt better and reject more—or if this level of churn is the new normal. The raw output is higher, but is the quality of the final artifact any better?
been there, migrated that