The representation of the problem may be difficult. You need to identify which variables are suitable to be treated as genes and which variables must be let outside the GA process.

The determination of the convenient parameters (population size, mutation rate) may be time consuming.

As in any optimization process, if you don’t take enough precautions, the algorithm may converge in a local minimum (or maximum).

For all the above reasons, if you know a traditional method to solve the problem, use it!

Try GA's if the system is partially unknown, non-lineal, or noisy. Even in this case, be sure that you are able to program the fitness evaluation function. In many cases this is by far the hardest part of the work.

On the other hand, these concepts have been very useful to provide solutions where traditional methods failed.

As a parallel computation process, the GA may explore the solution space in many directions and from many points.