Because a draft without context needs rewriting. It reads well, gets the customer's situation wrong, and correcting it takes longer than starting fresh.
The useful part of a reply is the specific detail — what they bought, what is overdue, what was promised — which is exactly what a general model does not know.
The draft is written from that customer's record: what they bought in Flow, what is being delivered in Loop, what is outstanding in Books, and their ticket history in Desk.
Every specific claim links to its source, so checking is reading. The agent edits and sends.
Nothing sends itself. There is no auto-send setting, on the rule that where a customer would find the error, somebody should see it first.
It also will not commit to a refund, a credit or a date.
The five-minute reply becomes a thirty-second one, and the follow-up nobody had time for actually goes out.
Consistency improves, because the facts come from the record rather than from each agent's recollection.
Yes, from the same sources.
It drafts nothing for that point rather than inventing it.
It weights the phrasings and sources your team keeps.
We will run it end to end on your own numbers in half an hour, and tell you honestly which parts Treepie does not improve.