AI support assistants: where people belong

Finding instructions, explaining a status or collecting an enquiry are useful starting points. Separate informational answers from actions that change a system. Explaining a refund policy is different from authorising a refund.

Describe supported questions and escalation conditions. Make the assistant’s role and the route to a human clear.

Instructions need owners, review dates and a clear scope. Connecting a model does not resolve contradictions between documents. Remove duplicates and separate internal notes from customer-facing material.

Retrieve order-specific information from an authorised system instead of inferring it from conversation. A missing source should lead to a clear limitation, not a guessed answer.

Include the question, attempted steps, relevant documents and the unresolved issue. Customers should not have to repeat a conversation.

Provide an obvious human-help action and an honest waiting state. Outside operating hours, offer a real continuation path rather than implying an instant response.

Build a set of common questions, ambiguous requests and cases the knowledge base cannot answer. Review correctness, source support and escalation after changes.

Automation rate can rise while resolution quality falls. Track task completion, repeat contacts and operator corrections. Confirm that private records remain inaccessible and state-changing actions require appropriate authority.