I don't see that as being an important objection, because from the point of view of any single human… that's already true for the expertise found only in other humans.

I don't know the chemistry necessary to turn crude oil into any plastic, I don't know how to use a Lagrangian, I tried and failed to learn group theory (I've only got the basics, and I'm not confident about them), and I've still only got a toy model of special relativity (which is supposed to be the easier one) — and yet, not only are those all still useful, the four colour theorem is useful even though it's a non-surveyable proof: https://en.wikipedia.org/wiki/Non-surveyable_proof

Depends on whether we can instruct the AI to make something useful out of it. Today we have complex systems like CPUs that most humans don't understand.

Abstractions are a thing. We use devices all the time that we don't understand entirely the workings of.

Or an answer so simple: 42.