Language models now produce software at a pace that was unthinkable until recently. Expert circles are therefore debating a side effect: a growing share of code that nobody in operations fully understands. Research paints a mixed picture. Gains are substantial for prototypes, yet in mature systems studies have even measured slowdowns, and without clear guidance generated code frequently contains vulnerabilities. Our conclusion is simple: the technology shifts the bottleneck from writing code to judging it. What organisations need are people who understand architecture, data models and operations, and who can assess what the machine delivers. In the ERP context we recommend firm guardrails: production code only with expert review, extensions cleanly separated from the standard, automated testing as a duty, and priority for proven standard processes. This preserves the speed advantage without sacrificing quality and maintainability.