As for moving to this versus your bespoke feature extractor, _given_ that your existing features and their compositions are not antithetical to their language representation, an LLM will be an equal or better feature extractor. But for example, considering the basic feature genre, if your collaborative filter has learnt features that smartly recommending a show with one text feature "sci-fi" to people whose preferences have the text feature "comedy" because of learned behaviour despite the text, then you have to verbalise this feature "scifi,laugh track" or maybe providing samples of the subtitles of the show or add a "frequently co-watched with" section (which contains comedy shows) to the prompt, to effectively get an LLM to do the same thing (or many other ways to induce a hybrid embedding)
In many cases, people have almost entirely verbalizable features and feature compositions in their existing systems even if it may not exactly be optimal. So it's a good idea to try out LLMs there.
Composition mentioned everywhere above is crucial. Provided you can verbalise your important features, LLMs can perform very strong deductions and compositions out of the box above and beyond our own feature interactions that we use with say xgboost setups. And it's dynamic in the sense that it gives you a foundation model you don't have to retrain to use new (verbalizable) interactions.