It feels like a premature optimization thing. He's got it in his head that the Labs are going to eventually be so prolific in their output that its going to be an undue burden for a team member to sit down and record a script for a labs result, and they don't want to deal with the risk of making the AI switch later and losing audience to the change, so just AI right out of the gate, what could go wrong. When realistically, one of the major elements of any well balanced work schedule is low effort tasks. You want to have low hanging fruit for people who are a bit burnt out, for the twilight hours at the end of the week, or just for doing if you end up in a situation where all your assigned work is running and your just waiting for results. Lots of situations in which someone could take a jaunt over to a recording booth, grab a test result script, spend an hour recording a few takes, then come back to check in on whatever stress test they were running.
Sure, theoretically you could hyper optimize to the point that nobody has that downtime, but now you're AI 'optimization' requires another razor edge optimization to even make sense, and both of these are just reducing the quality of the workplace and the quality of the product, making something nobody wants. And then as you say, modern AI is fairly error prone, and its not real time, so each retake is longer than a human retake. Tardwrangling them likely takes longer, and its not going to get any faster. At a minimum, you have to listen to the entire clip, at normal speed, to make sure there's no pitch or tone errors, artifacts, etc that might be masked at higher speeds, so your bare minimum overhead is the entire time to record it yourself anyway. Any failure rate greater than 0% means its more expensive in a time sense.