I Went All in on AI. The MIT Study Was Right (leadershiplighthouse.substack.com)

Josh Anderson spent three months doing what many companies dream of—going 100% all-in on an AI coding tool (Claude Code) and having the AI write every line. The product shipped and early productivity metrics looked great, but when a small change was needed he realized he no longer trusted or understood the codebase. Despite 25 years of experience, he became a “passenger”: unable to debug, maintain, or justify architectural and product decisions because he’d abdicated ownership to the model. He frames this as living proof of the MIT finding that ~95% of corporate AI initiatives fail—not from bad tech but from misplaced reliance. The technical implication is clear: AI can accelerate development but it can also hollow out critical skills—debugging, architectural reasoning, maintainability and decision ownership—leading to brittle systems that teams can’t evolve. Anderson proposes a simple principle: AI + HI, where Human Intelligence must dominate. Organizations should audit whether people can explain decisions, fix breaks without AI, and are growing their craft rather than their prompting skills. The winners will be teams that use AI as an amplifier while preserving human oversight, mentorship, and institutional knowledge, not as an autopilot that creates short-term gains and long-term risk.