Process Discipline as the Isolating Variable in AI-Assisted Enterprise Software Development: A Natural Experiment

Over 12 months (March 2025-March 2026), a two-person team using AI coding assistants produced 2.6 million lines of code across 2,929 commits and retained approximately 250,000 lines: a 90% elimination rate. Every line that survived was reviewed, evaluated against enterprise requirements, and kept or discarded by a human architect. GitClear (2025, 211M lines across Google, Microsoft, and Meta) reports that industry cleanup activity collapsed from 25% to under 10% after AI adoption.

This team sustained 90%. The dataset spans two process conditions applied to the same team using the same tools. Under a structured AI-specific SDLC: 1,484,038 lines produced, 811,684 lines eliminated (55% cleanup ratio), 3,977 function points delivered, 18 of 18 enterprise audit dimensions satisfied, 68 test files, 13 CI/CD pipelines. Without it: 1,153,069 lines produced, 483,991 lines eliminated (42% cleanup ratio), 428 functional FP delivered (working demo with live prospect accounts), 2 of 18 dimensions satisfied (including plain text password storage), zero test files, zero pipelines, four repositories abandoned.

The unstructured condition generated code 4.2x faster by volume (360,334 lines/month vs. 84,802) and produced working software that cannot be sold to a regulated customer and must be discarded rather than remediated: zero enterprise-ready function points. The structured condition produced at 9–11x the published elite productivity benchmark (227 FP/FTE-month vs.

20–26 for best-in-class manual teams; Capers Jones/SPR, 26,000+ projects).