We sell the AI. We would still rather you left it off for the first month, and this is the honest reason why.
An AI that answers support tickets is only worth paying for when the same question keeps arriving. If every ticket on your server is about one person's account, one person's payment, one person's ban, then a model has nothing to generalise from, and what you get back is a well-written paragraph that resolves nothing and annoys somebody who was already annoyed.
So the question is not "do I want AI?". It is "does what people ask me repeat?", and you can answer that by reading your own tickets.
The signal that says yes
Open your analytics after a couple of weeks and look at what the tickets are about, not how many there were.
If more than about a third of them are the same handful of subjects, like how do I link my account, where is my refund, why can I not join, what are the rules for X, that is a server where answering automatically pays for itself. Those questions have one correct answer, it does not change week to week, and a human typing it for the ninetieth time is a human who has stopped enjoying support.
The signal that says no
If most of your tickets need somebody to look something up, whether that is an order, a ban, a payment or an account, the AI is the wrong tool, and it is wrong in an expensive way. It cannot see the thing it would need to see, so it either says it does not know (fine, but you paid for that) or it writes something plausible-sounding that a member believes (not fine).
There is no shame in a server where the answer is no. Plenty of support is genuinely case by case, and a ticket bot with no AI in it is still the product.
The step almost nobody takes
You do not have to turn it on for everything.
Switch it on for one category, the FAQ one, or the one whose answers you are tired of typing, and leave the rest alone. You get a real measurement instead of a guess, on the part of your support most likely to work, and the categories where a human has to look something up never wake the model at all.
If it works there, widen it. If it does not, you learned that for one category's worth of credits.
What to write before you switch it on
The AI answers from your documentation and nothing else, so the documentation is the whole job. And the thing people get wrong here is the direction: they write more.
A twelve-line FAQ beats a forty-page manual. A short document that answers the question directly gives a precise answer; a long one gives the model more places to find something adjacent and more room to reason its way somewhere you did not intend. Write the answers to the six questions you actually get. Add the seventh when it arrives.
And the part that matters most
Whenever you switch it on, it stops the moment a person picks the ticket up. One press of Claim and the AI is out of that channel: no competing replies, no member reading two different answers, nothing to undo.
That is the bit worth designing around, and it is why turning it on later is cheap: the bot you have today does not change. The AI is a layer over it, and the layer comes off.