Some studios have already tried this. See Where Winds Meet for one of the most high-profile recent examples. I haven’t played the game myself, but my understanding is that the results are… weird. Not surprisingly, users pretty much immediately figured out how to coax unintended, game-breaking behaviors out of the AI NPCs.
But silly bugs aside, I think the main issue here is cost. So far we’re only seeing features like this in games with aggressive monetization, and that makes sense. LLMs are expensive to run. Getting good voice actors isn’t cheap either, but that is usually a fixed cost; you pay them once and that’s it. With AI, you’re paying for every single line of dialog uttered for as long as your game exists.
There are also no-zero setup and maintenance costs where you have to design specific guardrails to keep the AI from acting out of bounds. “Don’t give the player free loot, don’t use profanity or slurs, don’t discuss politics or sensitive topics with the player” et cetera. Of course players will always find ways around that, so now you’re playing a constant game of whack-a-mole trying to get this thing to behave the way you want. You’ve created a situation where you’re constantly paying for costly AI compute and you have to keep an “AI whisperer” on payroll. Suddenly paying a VA doesn’t seem so bad.