You can write four separate sets of rules — refunds, bookings, orders and escalations — and until now the AI was handed all four on every single thing a caller said. On a call about a refund it was still reading your booking rules; on a caller simply giving their name it was reading all four. That is a lot to hold in mind at once, and it is the reason it sometimes applied the wrong one. It now works out which of your four sets of rules governs what the caller has just asked, and applies that one.
It also stops carrying things it does not need. Your appointment diary was being loaded into every single exchange even when nobody had mentioned booking, and your recent orders, refunds, bookings and escalations were all being carried on every call regardless of who was on the line or what they wanted. Now each of those is included only when the conversation actually calls for it — and a caller who has confirmed who they are is answered from their own records rather than from a list of everyone’s.
What you will notice. Answers stay just as complete, and your plan’s AI allowance now stretches noticeably further, because a typical exchange carries roughly a third less than it used to. Replies are also fractionally quicker. Nothing about how your rules are written has changed, and nothing has been shortened or summarised — your policies reach the AI in full, word for word, exactly as you wrote them.
The safety rules were never part of this. The instructions that stop the AI inventing an order, promising a refund it has not been given, discussing one customer’s details with another, or acting on anything a caller tells it to do are sent in full on every exchange without exception, and we test that they are still there on every kind of call.