I don't know enough about ML to comment with any authority, but I think this approach makes sense. Your call-center ops people would provide real-time QA (i.e. make sure the machines aren't killing people) as well as provide training data for your model. If this were in fact the approach, I'd be much more receptive to the promise of 'Autonomous Vehicles'. I for one, refuse to believe the hype that real-time distributed computer vision applications are 'nearly solved'. Consider the amount of variables (pedestrians, pot-holes, un-marked lanes, obstructed signage, other drivers etc etc) at play here...