Start with a bounded task. Write recovery paths for unexpected input and preserve useful context during handover. AI features need verified knowledge sources, scope limits and ongoing review.
Conversation design
Tell people what the assistant can do and make escape or human-help choices easy to find. Ask only for information required to complete the task. Confirm consequential choices before acting and avoid trapping a customer in repeated recovery prompts.
Operational review
Test ambiguous, abusive, empty and out-of-scope input. Review failed journeys, handover quality and knowledge freshness on a schedule. Sensitive or regulated advice requires additional business and legal controls.