Yes. And no.

Ai can be more limited and setup for more specific purposes. This is possible with the current GPT-3 architecture with decent results.

But there are two problems:

First is that it needs to be trained and designed in a way that supports this more limited scope, which given most games is beyond the capability of the lore to have enough material to train an ai without the character getting access to information that they shouldn't know while still having enough material to make it coherent without just quoting. And yes, this does include Elder Scrolls. Even with the material, this is a long and expensive process, and usually easier to just code the dialogues you want that they want. Have to remember, it isn't just the AI outputting text, but also understanding players that may not always have the correct vernacular or method of talking.

Second, either the ai needs to connect to a service, which can lead to a variety of issues, or it needs to be run locally. With it being run locally, there are significant hardware requirements that make even a simpler model AI more demanding than what most gamers have access to. To add both the AI and the game in the same computer would make the demands even higher. So, running it as a service is the more viable option. But as a server side thing it has a higher delay, is more subject to moderation, in addition to being a continual cost for the company. Which doesn't bode well when most games as a service don't last a year.

But this is all subject to change. A more specialized ai could have lower requirements and be designed in a way to have both a larger language model with more limited knowledge libraries instead of a singular general knowledge set. A company could develop such a lightweight AI to multiple developers to make it cheaper and more unified.