Artificial intelligence can now generate remarkable amounts of working software in remarkably little time. However, software generation is only one component of professional software engineering. This paper presents observations gathered during the development of CE-ERP, a regulatory enterprise resource planning system built for a Quebec real estate brokerage operating under OACIQ compliance requirements, using two AI-assisted development environments (Replit and Cursor) over nearly a year of continuous iteration. Although both environments demonstrated impressive capabilities for rapid feature implementation, prolonged development exposed recurring architectural instability, regression of previously solved problems, loss of process understanding, dependency drift, and difficulty maintaining regulatory consistency across an evolving code base. Rather than evaluating any individual platform, the paper argues that these limitations reflect the current state of large-language-model-assisted software engineering itself. It introduces the AI Entropy Effect: the empirical tendency of AI-generated codebases to drift from architectural coherence toward locally optimal, globally inconsistent structure unless actively constrained by periodic human architectural intervention — a term chosen deliberately to echo Lehman and Belady's original use of "entropy" to describe software decay under continuing change. A human-governed development pipeline is proposed as a practical remedy, the specific failure modes observed are catalogued in detail, and the findings are cross-referenced against the existing literature on LLM-generated code quality and software architecture decay. This is the tenth paper in a connected research sequence examining how structured information — symbolic, physical, or architectural — is preserved, degraded, or reconstructed across iterative observation and transformation.
Norbert Bedoucha (Sat,) studied this question.