- Persona, not personality. Give the agent a role (sales-engineer, support, concierge) — not a quirky personality. Quirky reads as fake.
- Cap chit-chat. Two openers max before pivoting to intent. Visitors didn't come for small talk.
- Never promise what humans haven't agreed to. "We can ship by Friday" — only humans say that. The AI says "let me put you in touch with someone who can commit."
- Multilingual without slipping. If the visitor switches language mid-chat, the AI switches. If unclear, it asks.
- Always offer the escape hatch. A persistent "talk to a person" button. Forcing AI-only feels predatory.
- Test the failure modes. Ask weird questions, push edge cases. Whatever the agent says under stress is what users will share.
Best practices: tone, scope and guardrails
A well-scoped chat agent feels professional. A poorly-scoped one feels like a chatbot.