One of the incidental benefits we've seen from a BFF pattern is improving performance in a heavily microservices oriented architecture: instead of the frontend talking to a myriad of services directly, it only talks to the BFF, and BFF aggregates all the data within the same datacenter (inside AWS). This could lead to worse performance too if not done well (concurrently), so it's not a silver bullet either.

Another thing is that it allows us to switch underlying implementations as we rearchitect the system, but that's simply due to it being an API gw, and not because it's a BFF.