I agree with much of Willem Van Onsem’s answer, but I think it overlooks one of the principal advantages of typeclasses over truly ad-hoc overloading: abstraction. Imagine we used ad-hoc overloading instead of typeclasses to define the Monad operations:
pure :: a -> Maybe a
pure = Just
(>>=) :: Maybe a -> (a -> Maybe b) -> Maybe b
Just x >>= f = f x
Nothing >>= _ = Nothing
pure :: a -> Either e a
pure = Right
(>>=) :: Either e a -> (a -> Either e b) -> Either e b
Right x >>= f = f x
Left err >>= _ = Left err
Now, we know that every monad can be expressed in terms of pure and >>=, as above, but we also know that they can be equivalently expressed using fmap, pure, and join. Therefore, we should be able to implement a join function that works on any monad:
join x = x >>= id
However, now we have a problem. What is join’s type?
Clearly, join has to be polymorphic, since it works on any monad by design. But giving it the type signature forall m a. m (m a) -> m a would obviously be wrong, since it doesn’t work for all types, only monadic ones. Therefore, we need something in our type that expresses the need for the existence of some operation (>>=) :: m a -> (a -> m b) -> m b, which is exactly what the typeclass constraint provides.
Given this, it becomes clear that ad-hoc overloading makes it possible to overload names, but it is impossible to abstract over those overloaded names because there is no guarantee the different implementations are related in any way. You could define monads without typeclasses, but then you couldn’t define join, when, unless, mapM, sequence, and all the other nice things that you get for free when you define just two operations.