I once hacked in an A* implementation into Jedi Knight: Dark Force II. The in built AI in the game was very primitive (as most AI was at the time) but cog language provided me a way to work around it.
I initially used hand placed objects for the path nodes but I was experimenting with having the game rain down falling objects and keeping track of where they hit the ground to create a set of path nodes automatically. At first I was just going to look for negative space with a cog verb and set the nodes just above, but this doesn't catch walkable surfaces that are 3D objects instead of level surfaces. I then had to rain a node down from the stop point to get walkable areas beneath them.
There were still entity specific concerns that were not fully worked out. Entities had different sizes so could fit in different spaces. I had accounted for crouching and set the nodes to be more expensive when that was required as the entity moved slower. I had max jump distance somehow factored in and just had enemies detect a pit dynamically and jump when needed, although this made it easy for me as a player to block their landing site and force them into a pit.
Doors were interesting too, I never really solved that. My idea was to have a second metadata factor beyond node distance that would make paths through doors very expensive when they were closed so the entities would find away around if possible but hang out at the door if they weren't. You'd still need the door to manually communicate with the nodes since some doors are lockable but many just open automatically so are not an obstacle. That required a bunch of manual work for the level designer which I didn't like.
It was a very fun set of problems to work through since the base problem is simple (in that it is solved) but there were lots of interesting edge cases that weren't. And trying to make it work in a dynamic changing environment really changes the whole problem.