I managed an AI engineering team with HITL (in this case WITL) in place and my experience mirrors the article. A worthy read for anyone considering entering the space.
We all want the efficiency of AI, but it does take a lot of human effort to achieve the required accuracy first.
This is true of human intelligences as well. Faster than ML but doesn't really scale. HITL is a way to make it scale.
This is right on; but it is still difficult to pull off in practice. I expect that 5-10 years from now it will be somewhat standard/turnkey.