Choose the questions the assistant should handle, its sources and the cases it must escalate. Keep business information current and distinguish drafting an answer from taking an authorised action in another system.
Scope website, WhatsApp and voice channels against provider approval, account facilities, usage charges and privacy requirements. Define which conversations staff can access and when automated replies should stop.
Test representative questions, missing information, handover and provider failures before launch. AI responses can be wrong; agree human review, monitoring and correction processes rather than treating automation as a guarantee of accuracy.