I disagree, this is actually how you do it. You have the same AI for all difficulties, but you just disable some branches of the behaviours for lower dificulties. That is, in higher difficulties the AI gets to use more advanced behaviours. But you not using a different AI, you are using the same one and you only need to QA the specific branches that only advanced AI uses. This is not so hard to implement, and a lot of games use already this strategy. You can add AI bonus on top of that, and even just increase chances of some agressive behaviors or some counters to player strategies. And there you got it, an AI that gets better with higher difficulties and with only a fraction of the cost of a new AI.