So I feel like you're right in that we can do a better job of explaining how to do this (I obviously work at Wolfram Research).

But I think it's very possible and a strength of ours. Wolfram|Alpha is used by many services, including now MS Excel which I think is a counter example to the paradox you mentioned.

If you're interested in NLP applications, you can build APIs in WolframLanguage using the Interpreter (https://reference.wolfram.com/language/ref/Interpreter.html) or PLI framework (https://reference.wolfram.com/language/ref/GrammarRules.html) that give a finer grain control that doesn't require the full, overwhelming "full smartness" of Wolfram|Alpha. They're both pretty unique technologies and I think that also has affected adoption.