Okay, the simple answer is that the AI never simulates the outcome of its decisions. In general, most AI routines are based on some kind of predefined heuristic. This heuristic takes the current value into account, but it needs to be coded to be aware of what properties exist. For example, the civic AI considers the value of all civic effects from BtS based on what they are. So if you give a civic +10 commerce per Town it knows both that the civic is very valuable and knows that Towns are more valuable when it has the civic (because extra commerce for improvements is a concept that already existed in BtS and is already accounted for). This value however is never a simulation but more like an estimate that may or may not simplify the current situation and ignore certain aspects.
If I add a new property to a building (e.g. the Empire State Building effect does not exist in any way for any BtS building), the AI will not know it is there, and assign no value to the building. I will either have to add AI code to make it understand the effect, or apply a flat AI value to the building to give it some value.
So the general answer is that if I change a particular building, the AI will adjust its evaluation to the best of its ability (of course if the AI code was already bad that may or may not have the desired result). If I change what buildings can do overall, I need to teach the AI how to evaluate this new thing.