A study of more than 100,000 GitHub developers, published in May 2026 by MIT and Wharton researchers, put a number on a failure I've watched in every enterprise AI coding rollout: conversational agents drove a 741% increase in lines of code written — and a 20% increase in software actually shipped.
We measured activity and called it productivity
That 741% surge attenuated at every downstream gate: 65% more code reached review, then 20% more releases. The authors call it the weak-link hypothesis. AI speeds up the one stage it touches — code generation — while review, integration, and release stay human and fixed-capacity. You didn't remove the bottleneck. You moved it downstream and fed it harder.
I've seen the dashboard version of this failure: leaders celebrating record commits and PR volume while cycle time and release cadence flatlined. Throughput of an unconstrained stage isn't throughput of the system.
The perception gap is the governance problem
METR's randomized trial makes it harder to ignore. Sixteen experienced open-source developers, in their own million-line repos, predicted AI would make them 24% faster. They were 19% slower — then, holding their own timing data, still estimated 20% faster. A 39-point gap between felt and actual performance. METR's follow-up concedes newer tools help, but developers keep overestimating their gains by 40-plus points. And a Coddy Tech survey found 80% of developers call AI coding more dependence than advantage.
That's why rollouts succeed on paper and fail in production. Fluent output reads as speed; volume reads as value. Neither is.
The fix is instrumentation, not more tooling
Stop rewarding lines, commits, and PRs. Instrument release cadence, lead time, and review wait time — then staff the new constraint, review, before buying the next agent. If your AI coding program's only evidence is more code, you don't have a productivity program. You have a traffic jam wearing a progress bar.