I find it amusing that they tried to automate it with variables, etc, but seemingly wanted to make LLM generate a new response every time. Standard once written or even generated rejection form could’ve been more economically sound. But, but - then you wouldn’t up the metrics of AI usage in your department!