He's talking about the value proposition RIGHT NOW which is for higher accuracy but also more expensive (because you'd need the same human in the loop + pay for the AI) and how that's different from how AI is being pitched right now which is save money and use less people. He's not assuming it won't get better, he's saying it's what it is right now.
We've been doing that sort of AI + radiologist since the mid 90s now, and it has made an incremental improvement; I expect it will continue to make incremental improvements. It's certainly cheaper than double reading everything (which has been show to improve some things, e.g. screening). Anyone expecting a radical change in that model in the short term is unfamilar with the SOTA in one or both domains.
To put it another way, if we are looking for places to get a human out of the loop, radiology shouldn't be that high up anyone's list - but decision support is super useful sometimes.
