I think in a lot of cases that's because the meta with LLMs right now is to "have them do things for you", which generally means that obfuscating what's actually happening behind the scenes can make them seem "smarter" to the average user. Also, engineers are used to full control over deterministic input-output pipelines, which is a framework they try to force on LLM applications that fails miserably for the reasons you've listed.

In my opinion, the best applications of LLM UX will have full clarity for the end user (something we're trying to do with MatterRank). The non-determinism should be something the user can control to get better results, not something the engineer has prompted that takes control away from the user.

Now, if the use case you're looking for is "give me results with x text", then yes I agree with you that LLMs are just getting in the way. But that's not always the case.