[jtheory]

It's a different type of error that's possible, though.

Manual entry errors happen at predictable rates, and with good processes you can get them down quite low.

Software error rates are not so predictable (unless you have full control over all variables... and that's surprisingly hard to get). You could deploy this software in the context of a slightly different version of the radiology software which used a different font in PDFs (or whatever) and find that all 8's are now read as 3's; no human would do that.

If a human started getting input that wasn't sufficiently readable, they'd talk to someone about it. Could the software do that?