Not just a level of indirection. The "substantial features" of the code need to not be directly exposed.
So if you had some AGPL OCR tool you were using, you could use it, but not in a way the user sees that text. Generate audio from it and expose the sound? Probably fine.
My understanding of the theory that they're advocating is that the AGPL requires that when you modify AGPL software you modify it to provide an offer of its source.
And that you can comply with that completely, run the software, and then have a proxy in front that strips that offer without violating the letter of the license.
And if that theory works I think "substantial features" of the code could be directly (but for the indirection of that proxy) exposed.