Are you thinking of Erlang? Erlang is definitely not a ML, although they share some similarities (immutability, pattern matching). Not sure what ML originally was designed for, other than maybe building compilers.
> So a person might choose to write a web application in F# because they are thinking about the problem in terms of a state machine
Algebraic data types in ML allow you to encode program state within the type system, so the compiler can help ensure the program is in a consistent state at any point during run-time. That is the real benefit to having a strongly static type system in the vein of ML.
Regarding the original question: the difference between static and dynamic types is, IMO, one of checking at compile-time vs run-time. (But the static/dynamic definition is fuzzy enough that there's often leeway for interpretation here.) Type inference is icing on the cake -- type annotations could be explicit (C, Java) or implicit (ML), but a language is not statically typed unless the types are known prior to execution.
ML stands for metalanguage. It was originally developed for a theorem prover. I think the expressibility of the ML family of languages turned out to lend themselves to many categories of programming problems like compilers. Especially since the code could so closely match the structure of the problem itself (via recursion, pattern matching and other features).
EDIT: Commenting on my point about code matching the problem structure. This is why I'm, in general, a fan of most functional and declarative programming languages I've tried. Prolog, Common Lisp, Scheme, SML, Ocaml, Haskell, Erlang, etc. They may not be ideal for every problem (though the malleability of Scheme and CL make them pretty damn close, IMHO), but there are a number of domains where they truly excel compared to the bog-standard enterprise languages. The only problem I have not solved is convincing bosses to accept a polyglot environment. They seem happy with Python and C++ and C# and Java, but throw in even F# and they reject it.
Static typing offers some advantages in regard to tooling. Entire classes of errors can be caught by the editor/IDE as the programmer types code without waiting for compile time...or in the case of dynamic languages perhaps runtime.
ML languages are particularly suited for the classes of problems for which they were developed, e.g. fault tolerant high reliability systems, because they can efficiently express state machines via pattern matching semantics and the SOME|NONE construct.
So a person might choose to write a web application in F# because they are thinking about the problem in terms of a state machine. With code generation, a stack similar to Microsoft's can provide a lot of the Presentation and Storage layers in Javascript and SQL respectively.