While I agree with most of what is written above, I would like to refine the "how to start" ideas a bit.
My suggestion to you is to never ever program directly to EJB interfaces within your code. Always use a regular, business-oriented interface, program to it (meaning, have your code call methods on the business-oriented interface) and provide the EJB "glue" code as a pluggable implementation. Your program should be focused on business logic, and not on implementation details such as EJB.
That way, you can easily switch between remote and local implementations - and if you use an IoC container such as Spring, you can do it by means of configuration only.
A special note about switching from local to remote: note that there are a few semantic differences between the two. For example, calling an EJB method via its "remote interface" results in arguments being passed by-value, while calling through the "local interface" results in arguments being passed by-reference. This is a major difference; so if you "start with local", make sure that you design your system in a way that it takes "remote" semantics into consideration as well.
If your design relies on EJB methods changing passed-in objects, then it would be tricky for you to "switch to remote" later; perhaps even impossible.