But I'm not sure some of the issues are really AI. For instance the SU can't defend against Germany but this isn't really an AI thing. I think it's the whole reserve unit MP concept and this is a major design change.
This is certainly a factor, but it is still primarily an AI problem, not a design problem. When the SU sends all its units into Siberia, leaving nothing in the west between hitler and Moscow, that is definitely an AI problem.
When I talk about AI, I am talking about much more than tactical military AI (this is actually the AI that currently is in the best shape).
Strategic / theatre military AI should know to not leave an entire front or coast undefended and also to have some sort of reserve forces to respond in case of a surprise declaration of war. It should also know how to transfer units from one theatre or another to adapt to a situation. It should know how to set high level objectives for itself (invade germany, hold the western front, win the convoy interdiction war, etc) and then influence the production and research AIs to help meet those goals, while the tactical AI figures out how to win any battles needed to reach those goals, while sensibly coordinating among the different military branches (army, navy, air) to meet those goals.
Production AI is where you could maybe teach the AI that building only reserve units is maybe a bad idea, no? As well as teach the production AI to do manpower checks based on how many reserve units exist so it will at least have sufficient manpower to mobilize if needed. Also you could teach it that in an emergency mobilization situation, it should prioritize reinforcements rather than unit production. I would bet any money that it is currently the opposite, with unit production staying fully funded and reinforcements getting the leftovers. This is the right move most of the time, but not in this situation.
Or you could teach the diplomatic AI to mobilize when certain threat of war escalates (hard to pin down the triggers for this) or generally to pick laws better (right now I believe it's just one line of code that says basically pick the most "advanced" law available).
Of course, research AI could maybe be taught that if you have a large reserve infantry army, some infantry techs are probably a good idea? And maybe some infantry doctrines, especially those that boost org and morale? And as Russia you probably have supply problems so would be nice if the AI could prioritize supply techs.
The list goes on and on. To conclude, the problems is not fundamentally a design and mechanics one (I actually think the overall design and concepts in the game are excellent, although these are certainly also areas that can be tweaked and better balanced), but rather the AI often has no clue how to use the existing design, features and mechanics. Not to mention we are really talking about over a half dozen different AIs (theatre, HQ/Army/Corps AI, Production, Research, Intelligence, Diplomacy, Trade, probably some others I missed) that have absolutely no idea how to coordinate together to create a comprehensive, consistent, synergistic overall strategy.
Hence my naive and no doubt incorrect hope that Paradox will pour most of their development resources with SF into the AI, not into adding in even more new features.
Already, with this new achievement system, it begs the question, will the strategic AI have any idea how to use it? Will it actually try to set itself up to meet the criteria to unlock any of these achievements? And will the achievements it decides to pursue make any sense, fit in at all with the overall geo-political, economic and military strengths of the nation? Or will it simply be achievements the AI accidently unlocks by happening to coincidentally meet all the criteria? My guess is the latter.