I think the real problem the AI faces is a buggy payback calculation of troop disbanding.

It counts like this:
if i disband this unit, i get 300 per turn, but in order to recruit it i need 1170, therefore, I effectively break even at turn 4. What is the % probability i would need the unit before then probablility=function(ongoing wars, unit location, enemy closeby, ability to transport the unit to required spots, population control requirements, etc). It is at war all the time, the function calc is improper, so it gets that it does not pay to disband a unit. it doesn't.

a human player thinks more like: if i disband this unit, i will cover what it does now with those other units I have, hence, risk is increased only marginally (e.g. i can keep both the rebellion down and defend against enemies with a single stack in a region, no need to sit in the city). I will probably need to recruit a unit soon, but i can invest the cash flow now, so I get paid back all throughout the campaign through wealth effect (or exports if it is a plantation.). moreover, I am not sure i will need the new unit in this precise spot as I have several hot zones. therefore, while money is more mobile than units (strategic benefit, hard to evaluate for the ai), I'd rather take the money