Being able to filter, sort, or transform properties of the objects passed over makes it much more straightforward to work with than bash et al.
The memory overhead with sets of objects and their properties vs vanilla strings is significant though and easy to bump up against when working with large data sets. Best to try to keep all of the processing for that dataset in a single pipeline if you can. But that’s tradeoffs.
I had an idea where I'd make wrappers for all the popular commands that would accept JSON structured data instead of raw binary data and "do the right thing". And that way you could take advantage of it without having to change your shell tooling. And they'd all accept an extra argument called either --structured-out or --unstructured-out which would either emit JSON or render the output back to a "flat" string, as appropriate.