Let me start by saying that I don't expect the following to be implemented however I thought this would be a really interesting post with an idea I had that could make the AI much harder by making the ai smarter instead of buffing their stats.
I was reading a paper,
Paper , and started looking up examples of their presented 'NEAT' algorithm and stumbles across this video:
NEAT in action.
What I found interesting was the graph presented in this video at 3:52,
3:52. Showing the plateaux in which the AI struggled with. The reason I find this interesting is because with the right 'secondary' logic, (for example the logic of the current AI could be used as a fall back), one could implement these generations as set difficulties
Now, obviously the complexity of DOS2 combat is going to make running a simulation for any amount of time just impractical BUT here's where another interesting idea comes in: What if you ask your players to submit anonymous combat logs? All this would require is a decent structure for a machine to understand the difference between good decisions and bad ones and a method of analyzing and evolving all that data into a single test machine.
If done right you could end up with a truly difficult mode done right. Not only that but once the infrastructure of a project like this has been build it can be used for other projects and games.
Food for thought