In fact - FAANG companies have the resources to make horribly inefficient processes work (in human time or computer time) - I suspect many of the AI powered things are examples of this - and a smaller company will die trying to get it to work.
I don't think ML-powered systems are inefficient[0], so much as they are incremental gains over the pre-existing system that are only justified by huge sales volumes. If you can get a 4% lift in sales by using a complicated ML model with 100k features instead of a handful of basic heuristics, then whether it's worth spending engineering effort on depends on what your sales are.
[0] Inefficient in the sense that they're wasting money which could be easily reclaimed - they're probably going to be improved over time as the state of the art improves, but that relies on the whole field moving forward.
reply