The problem is that the "human" cannot accurately express himself in English, or cannot define the whole problem scope. Mastery of the English (or whatever) language that the Enterprise is using is more important than technical expertise. This cannot be stressed enough. Correct and extensive technical specifications are as hard as the code/stack itself.

The human here is acting as a compiler for the real world. In each iteration, it's raising exceptions and errors for the AI. The iterations will continue until the AI compiles the correct output.