Step 1: Define the problem.
Decide what _kind_ of AI you need. For instance, are you trying to build a rule-based engine? (Where will the predicate rules come from?)
Are you trying to simulate a neural net? Would fuzzy logic routines be more suited?
Are you trying to get an an answer, such as optimizing on a solution via hill climbing? Or are you trying to emulate higher level behavior and scripting might do better?
Are you trying to represent complex relationships? (Link lists/ LISP) Or are you trying to parse natural language?
...in short, figure out what kind of AI you're interested in, what problem your trying to solve, what solutions best fit that category of problem, and then focus on implementation language after the fact.