Why AI Coding Still Fails in Enterprise Teams (www.aviator.co)

Four senior practitioners — Kent Beck, Bryan Finster, Rahib Amin, and Punit Lad — argue that the viral demos and headline claims about AI writing large swaths of enterprise code mask a different reality: without systemic change, AI coding tools often slow teams, introduce technical debt, and fail to scale in large, security‑conscious codebases. They say enterprises repeatedly fall into the “shiny prototype” trap—deploying agents without training, spec discipline, or workflows—so generated code becomes fragile spaghetti that inheritors must fix. Cultural factors (fear of job loss), misaligned incentives, and the loss of tribal “why” knowledge further block reliable adoption. Technically, success depends less on model capability and more on context and process: spec‑driven development (refined, machine‑digestible specs), test‑driven workflows, prompt and artifact versioning, multiplayer prompt engineering, and audit trails. AI amplifies existing practices—good CI/CD, disciplined testing, and delivery pipelines get better; poor practices get worse. The implication for teams is clear: treat AI as a powerful tool that requires training, governance, and integration into existing software‑engineering practices (security, maintainability, delivery), not a drop‑in productivity panacea.