An FST allows you to accurately model 'outputs' on state transitions (where those outputs can be side-effects like 'make a call' or 'send an invoice.' This allows you to prove things about what sort of side-effects/outputs will necessarily have occurred by the time the FST is in a certain state.
I started with an FSM but, found in practice, I was using it like a poorly written FST.
Sometimes I used FSM's during the design phase of a project to model system behavior, but I'd end up implementing them pretty crudely with spaghetti code or case statements. Looking back, it would have made more sense to explicitly implement the FSM in the code.
Essentially my first take at the phone stuff I had done (a bit over 3 years ago) was to just roll into a single view and hack my way through it. The requirements (at the time) were relatively simple and this served us well for 3 years.
Now our requirements for phone calls has really changed as we've added things like voice prompts, whisper files, user interaction, and a bunch of other doo-dads. Worse yet, the presence (and behavior) of these states can now vary by the customer that is using them.
So I broke that jumble of code into a state machine with clear transitions. Each state farms the actual functionality to defined strategies which can be specified on a per-customer basis.
It works like a charm:)
After I got more comfortable with actually using them for real work (as opposed to "cute toy from CS class"), I started seeing FSMs everywhere I looked in my projects.
For one thing, they make analytics a lot easier to slot into your system. For example, I record the fact of interesting events like e.g. purchases and send it off to KissMetrics or internal analytics stuff for graphing/reporting/auditing/etc. That gets done in my controllers, and it results in a whole lot of duplicative code which I often forget to write. (It's easy to miss that e.g. if I credit somebody manually for a purchase that should count as a purchase, even though it is under admin functions rather than the purchasing pathway. After all, if I forget to write that log code, all that happens is my stats get silently borked.)
This could get fairly easily solved by having the logging code aware of the change in the user's status, which is about three lines if you have a state machine, and won't put the logging logic in 12 different places in your app or cause you to have to pollute the user model with a big ol' case statement of doom. Plus, if you want to log more stuff, you naturally go into the log logic and make it do more things when the right events bubble to it, rather than chasing down a) who "owns" the event or b) where the event happens.