HATEOAS Works with an LLM in the Mix
Most of the dev teams I've ever encountered who said they were "doing REST" were not actually following HATEOAS. Per a strict reading of Roy Fielding, he would consider that "not really REST." (Now don't get distracted, I don't want to wade into that whole purist debate).
The reason many did not do HATEOAS is that it requires the API client to be smart and adaptive. It would discover "ok, what can i do next", apply logic to it, and choose the next step. But many shops were on tight time commitments and it was much simpler to just think of REST as "json over http with consistent url patterns."
The cool thing is: With an LLM in the mix, HATEOAS is unchained. An LLM can do exactly what a "dumb" api client cannot: ask "what can i do next", and then use _inference_ to understand those options and select one.