I think there is a great opportunity in using GraphQL as a unifying entry point to legacy systems. In fact, when I talk to people at meetups and conferences, this is a topic that is almost always brought up. Companies that might have previously considered
https://apigee.com or one of their competitors now ask themselves whether they should rather adopt GraphQL instead.
When thinking about building a GraphQL entry point there are 3 main areas you will have to consider: Performance, Authorization and Complexity.
GraphQL is designed to be backed by arbitrary storage, so it would actually be feasible to have a bunch of GraphQL resolve functions that just call out to existing REST endpoints. To make this performant you need to implement a proper batching mechanism like DataLoader in node or Sangrias Deferred Resolver.
If your existing api has automatically generated docs in a format like swagger you might even be able to generate the entire GraphQL schema from this :-)