Objections are not surprises; they can be counted
Customer worries come down to a dozen or so: price, results, authenticity, timing, after-sales. Export three months of conversations and sort them, and the long tail turns out to be short. Anything countable can be prepared for.
Improvising is the expensive option
Letting the model improvise against an objection is a fresh gamble every time. The same question gets a different answer today and tomorrow, conversion cannot be attributed, and when something goes wrong there is no specific sentence to fix.
Turn frequent objections into reviewable rules
Each rule has three parts: when it applies, the facts that may be stated, and what must never be said. People write the rules, people review them, the system runs them. The model makes the wording natural; it does not decide the substance.
A model may choose the words. It should not choose the message.
Keep feeding the library from real conversations
Review the conversations that did not close, find the objections the library does not cover, and add them. After a few rounds, objection handling stops being personal experience and becomes something the team owns.