Edit (additionally):
If they can't be made atomic, make the transitions timed, with expiration, never allow deadlocks to occur and always branch to a fail-safe state. Players will tolerate things like the uncontrollable NPC getting pinched off screen and respawning, even with a comical animation if it's that kind of game; they hate a blind referee ruining things for no perceivable reason.
In the case of another game where someone's actually making progress up a very large structure, I can't think of any game offhand where saving in such a state has actually been allowed. However allowing a save there by necessity triggers the edited version of my statement which includes a timed state transition and safe aborts for failures in state change. Such a save would also include the current state of progression as well as the remaining limit for the timer.
Probably the most ludicrous is that when sneaking, only the player is taken into consideration by enemies. I've seen patrolling enemies kick my companions out of the way without actually noticing them.
Reading the player's intent would probably still be a very difficult problem, so even with the perfect knowledge of how to interact in the world they'd probably not act and react to changes in the player's intent anywhere near as well.
Skyrim was/is awful about that. I'd be sneaking along through a tomb, and then Lydia screams "I am sworn to carry your burdens!" and charges in, waking up a dozen draughir.
When large amounts of programming time is spent on game ai it does end up "self driving" at an adequate level, see Google's sc2 bot, Google's older arcade game bot's, and openai's dota bot. It's just not worth it investing that amount of time into most games.
They should probably also consider adding fall damage as part of the considerations in invulnerable mode.