I mean it definitely captures a relationship between parts of input data in ways that many other models cannot. It also cannot do everything.

Like most real world problems, there's a question of how you will represent the relevant information in data structures that best suit the ML methods you intend to use. Similarly there's a question of whether the ML methods will do what you want, given the data is in that form.

Despite the fact that transformers are killing it for ordered data, I'd say their flexibility to deal with unordered data is definitely of interest for real world problems where representations are tricky.