I understand the spirit, but that's how it goes. You have somebody doing the work, as you want the ML to do it, and then feed the data. It's the same when they get oncology scans that have been diagnosed by well paid doctors, somebody who knows does and the machine tries to replicate.

What very likely happened is that the failure rate platoed much higher than they expected, and all this time the goal was to lower it. Remember, it's cheaper to have 0 people in India than 1, specially with AWS in mind.

Moreover, even if the accuracy was incredibly high, they would still need people reviewing. You have to review random events to ensure the model keeps performing well and to evaluate the ones with low confidence or suspicious.