In this particular case, it seems mostly use cases that were solved some decades ago with _barcodes_, so it's not even cheaper or more profitable. There's already an alternative to human labour which works for better for this sort of application.

Computers can do some things better than humans; recognize and generate speech, when accents are involved at either end.

And nag and upsell well past the point humans feel comfortable doing.

Except there's still a human that has to confirm everything the computer did