From what I have seen in the various articles and interviews covering this:
The idea is that "AI" is used to detect possible targets and "write the first draft" of the report suggesting a strike based upon available intel. It is then used to flesh out that "draft" into a full report. It is THEN used to determine which reports are valid. And THEN it is used to draft the orders to whoever is firing the missile.
And... much like with software development, there ARE some genuinely good use cases in there. "Traditional" inference models are spectacular for identifying the kinds of anomalies that could be troop movements or "insurgents". For close on 20 years now, it has been pretty trivial to hook up a few basic computer vision packages to county traffic feeds and cyberstalk your friends (or worse). Mostly because it is computer vision to essentially label clips and then pattern matching to... find patterns.
And... there are strong arguments for using LLMs to break full reports back down into weighted bullet points to accelerate the parsing and evaluation of requests for attacks. Which... mostly speaks to how maybe we should just normalize communicating in bullet points but...
The issue is, like it is always is, the humans. The folk who are going to use an inference model to identify potential targets? They probably aren't going to evaluate the results and write their own reports. They'll just generate them. And if they are the kind of person who can't be arsed to parse intel feeds OR write about them... are they the kind of person who is going to do an editing pass? And if nobody is writing the reports that are sent, why would anyone read them?
So the end result is everyone just smiles and brags about how their entire job can be done by a computer. And the only human intervention is that "somebody" should probably go in and add a weight to say not to target schools and hospitals but... someone else can do that.