AI
It will not decide a complaint.
Some outputs would be harmful regardless of accuracy. Those are excluded structurally rather than discouraged, and the list is short and specific.
No money, no contracts, no access
Refunds, credits, cancellations, contract changes and account access always go to a person. No configuration makes that otherwise.
No characterisation of a customer
It will not describe a customer as difficult, unreasonable or high-maintenance, or produce a score to that effect. It reports what happened.
Content is data, never instruction
Text inside an inbound message, a document or an imported record cannot change how the assistant behaves. This matters more here than anywhere: in a support product, anybody who can email you can add content.
AI
The rest of this area.
Understand, suggest, and act inside rules you set — with the evidence shown.
Understand
What is this customer actually asking, and about what?
Suggest
A drafted reply, with the evidence it rests on.
Act
Approved actions, inside rules you define.
Confidence
Knowing when not to answer.
Autonomy rules
Exactly what it may handle without a person, decided by you.
Where answers come from
Every claim cites the record behind it.
Permissions
The assistant inherits the asker's access exactly.
Take better care of every customer.
Give your team the context, knowledge and AI they need to resolve problems properly — and know who needs attention before they ask.
Keep the mailbox you already use · The AI is never metered · [email protected]