Well, Modern neural networks and LLMs aren't constructed using any requirements, specifications or hard constraint at all - they just heuristically give a context-aware approximation of their massive input data set.

So getting to X being part of "requirements" is a bigger step than the layman imagines. Especially, 'cause an LLM outputs approximated recipes for action when asked, LLMs struggle to consistently achieve goals since goal-seeking requires constant little adjustments.

But even more, if an AI somehow reached the level where you could inject goals into it, it seems very likely it's goal would be maximize investor profits and this wouldn't solve the problem of the system discarding human experts but rather compound it.