Software needs to be transparent, a human expert worker needs to be able to verify the result a software generates within reasonable time.
There is no expertise for software that you could delegate to aside from an entity using some form of integration test.
This is becoming more and more important as the AI industry is pushing non-deterministic computing further and further. There will be court cases where the full automated decision chain will be called into question, and a default assumption of "well, we don't have any logs so we should assume the software did the right thing" terrifies me.
Even speaking from an IC/non-managerial place, far too many technical problems in businesses are really just symptom-alleviation (or worse, performative look-I'm-doing-something theater) for a deeper problem which involves misaligned incentives for humans.