If we're talking AI as the main issue to be addressed, then the real goal is not an NPC who can pass the Turing Test (which would be massive overkill) but simply one that creates the perception of plausibility within a combat context -- call it the Eliza Effect.
In this model, the goal is not for every NPC's every action to be correct, but merely to seem sufficiently correct for the situation. Which means we have to think about apparent intentionality -- what the NPC does has to appear to be aimed at satisfying what the NPC wants.
And how do we know what an NPC wants? She tells us! If I say, "Man, I'm thirsty," you won't be surprised when I go to the fridge and get something to drink. Your expectations of what I will/should do are conditioned by what I tell you about my needs and desires.
NPCs can benefit from this same effect. Just as people were conditioned to think that Eliza was a real psychologist, and then read into the program's comments what they expected a psychologist would ask them, players will be more likely to perceive as appropriate some NPC action that fits into what they've previously learned about who that NPC is and what he wants. So it seems to me that, from an AI perspective, developers can stack the deck by creating high-bandwidth channels for communicating information about an NPC to players before players interact directly with that NPC.
The masters of this kind of thing were the folks at the late Looking Glass, Ion Storm, and Irrational Games, and they used a very simple tool: vocalizations. NPCs would wander around, apparently following internal goals, and talking about those goals (or just mumbling out loud). When a player would get their attention, they'd produce some intermediate vocalization ("Hmm? Was that someone?"). And if the player then came clearly into their attention, they'd produce a definitive vocalization based on their level of hostility to the player ("Ah-ha! I've got you now, taffer!")
These vocalizations were brilliant (and nowhere more so than in the "Thief" games) because they communicated to the player something about the internal "motivation" of the NPC. So when the NPC then followed through with some action, it seemed appropriate because the player was already conditioned to expect that kind of behavior based on the information transmitted. Additionally, the clothes and gear of an NPC are also forms of information that are transmitted to a player that say something about who an NPC is, what their primary role is, and what they want.
So why can't MMORPGs do something like this? If an NPC looks up and yells, "Hey! Are you a medic? I hate medics!", then starts toward the group's healer, that accomplishes two things: it gives the group warning and lets its members react appropriately, and it reduces or even eliminates any surprise that the NPC goes after the healer of the group exclusively.
If all intelligent mobs (NPCs) are programmed with intentions, and are able to communicate those intentions through various visual and auditory channels, why again does that MMORPG need to copy the notion of "aggro" from other MMORPGs with a less well-developed system of perceived AI?