Loading
The repetitive majority: order status, plan details, password, where a setting lives. These are most of the volume and none of the difficulty.
What is not deflectable is anything ambiguous, emotional or commercially significant — and a bot that attempts those costs more trust than it saves time.
The assistant answers from the records and documents it is permitted to read, citing them. Guardrails define what it may assert and what it must hand over.
Handover triggers on low confidence, on request, and on any topic you have listed — and it carries the whole conversation and the customer record across.
It will not act on an account. Reading and drafting are safe; refunds, cancellations and changes go to a person.
And deflection measured alone rewards a bot that refuses to hand over. Read it beside satisfaction or it will mislead you.
Volume falls on the questions that were never interesting, and what reaches an agent arrives with context rather than as an opening question.
Everything the assistant handled is logged as AI-handled, so deflection is measured rather than assumed.
The repetitive tier answerable from your own records and knowledge.
Low confidence, on request, or any listed topic — with full context.
Resolution and satisfaction alongside deflection. Deflection alone is gameable.
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.