What a good answer looks like when information is missing
Say a customer asks whether they can pay for a course in instalments, and your documents say nothing about it. A good answer sounds something like this:
"I don't have that confirmed in our documents. My colleagues can answer you right away on WhatsApp."
An unsure "yes" or a vague reply that sets false expectations costs far more: an unhappy customer and, sometimes, a promise you can't keep.
What to prepare to reduce wrong answers
Answer quality depends more on your documents than on the assistant. Write down sensitive topics explicitly: exceptions, cancellation conditions, cases that need your approval. Whatever isn't written, the assistant shouldn't fill in from imagination.
"I don't know" doesn't mean "I can't help"
There are two kinds of questions. Those about your business (prices, dates, conditions) are answered only from your documents. General ones ("what's the difference between the two licence types?", "how do I prepare for my first course?") can get a useful answer from general knowledge.
A good assistant keeps them apart: it happily explains a general curiosity, but never turns a general explanation into your price or your rule.
A practical step for today
Pick 10 common pre-purchase questions and 10 post-purchase ones. Put each into one of three groups: clear answer in the documents, needs a fallback reply or should go to a person. That improves quality faster than trying to make the bot answer everything.