I'd rather say that STM is intended to be used just for building up transactions, not to model your whole process/thread, simply because in the latter case your process couldn't have any observable intermediate state, or put in another way, between any two transactions the information can only go in one direction.
So, since synchronizing two processes needs bidirectional communication, you have to use more than one transaction.
In the end you still write processes in IO, but communicate via transactions built in STM.

It's a good thing to expose your primitives as STM when you can, so that users can build larger transactions on top of them, but a synchronous send is not a transaction from the start so there's no choice.