It really depends on the type of AI that's implemented. The most common form is what's known as a Finite State Machine. Basically each player has a series of States they can be in, Transitions that define which States they can switch to, and Events which determine when these Transitions occur. A State could be something like "Get the ball", "Dribble the ball", "Pass the ball", "Get into range of ball" etc. Internally these would be programmed as behaviours that are switched between if certain variables (the aforementioned Events) are true. Scaling the AI is therefore a case of slowing down the decision making of the computer by implementing a time delay between Transitions.

That's the basic gist of it. I could go into more complicated detail.