Process Discipline as the Key Variable in Ai-Assisted Enterprise Software Development: a Natural Experiment

  Findings: Same team, same AI tools, different process conditions: 18/18 enterprise
  dimensions satisfied at 8-10x elite productivity benchmarks under a structured SDLC,  
  versus 2/18 dimensions with no tests, no quality pipelines, and no code review when 
  process authority shifted to non-technical stakeholders. Measured across 287 FTE-days,
   1.48 million lines added, 811,000 lines removed, and 18 enterprise dimensions derived
   from SOC 2, NIST, OWASP, DORA, CIS, and CNCF frameworks.                             
                                                           
  Scale: No comparable study in the AI-assisted development literature approaches this
  duration or granularity. Peng et al. (2023) measured a single isolated task. METR     
  (2025) tested 16 developers on individual issues. DeputyDev (2025) observed 300
  engineers but had no unstructured comparison arm. This study spans 13.7 FTE-months of 
  sustained enterprise development with commit-level traceability across multiple
  codebases.

  Literature gaps addressed: (1) No published study isolates process discipline as a    
  controlled variable in AI-assisted development. This paper presents a natural
  experiment holding team and tools constant while varying the development process      
  through an exogenous organizational change. (2) The speed literature produces
  contradictory findings (55.8% speedups vs. 19% slowdowns) with no reconciliation. This
   paper argues these are measurements of different process conditions, not conflicting
  results. (3) No published benchmark exists for sustained AI-assisted commit rates;
  this paper reports 10.7 commits/FTE-day over 287 FTE-days. (4) DORA's "amplifier"
  thesis rests on correlational survey data; this paper provides project-level evidence
  with a causal mechanism. (5) Model collapse research (Shumailov et al. 2024, Nature)
  has not been connected to practical codebase quality; this paper identifies the clean
  starting codebase as a multiplicative requirement grounded in generation loss theory.
  (6) The role of organizational governance in AI-assisted development quality has not
  been empirically demonstrated; this paper documents a quality collapse caused by an
  organizational decision, not a technical one.

  All metrics derived from git history, source code analysis, and published industry    
  benchmarks. Methodology described for replication.