That famous 55 percent AI speed-up has an asterisk
The famous Copilot speed study measured one bounded task. What controlled results do and do not tell you about your own team's experience.
The most quoted number in AI-assisted development remains GitHub's controlled experiment finding developers completed a task 55.8 percent faster with Copilot, per the company's published research. The study was real and reasonably run. The asterisk is what it measured: one bounded, well-specified task (an HTTP server in JavaScript) built from scratch by recruited developers.
Bounded greenfield tasks are the assistant's best surface: clear spec, no legacy context, no review step, no integration. A working week contains some of that and a great deal of the other thing, reading old code, negotiating interfaces, waiting on reviews, and the 2024 DORA findings suggest the system-level picture is more complicated than the task-level one. Both results can be true. Tasks got faster; systems did not automatically.
Our advice when a 55 percent lands in your planning meeting: accept the direction, reject the transplant. Measure your own before-and- after on the outcomes you already track, over a quarter, and expect a real but smaller number, concentrated in specific kinds of work. Vendors quote the lab. Your ledger lives in the field.