Not necessarily. NLP is possible to do without any ML. See ELIZA [0] as an example, which used pattern matching and substitution to process natural language. A few simple rules can cover a majority of requests. It's still AI, but not ML.

Though I do agree that the ML-based NLP thst we have today is much more usable, albeit more compute-expensive.

[0]: https://en.wikipedia.org/wiki/ELIZA