Support AI fits in-knowledge answers and pre-routing; promises, refunds, safety, and emotional escalation must go to humans. Policy beats model brand. Drive with an acceptance checklist, not slogan shopping.
- Build citable knowledge with expiry before wiring an LLM.
- Answers should carry sources or ticket IDs to limit hallucination damage.
- Escalate on sensitive terms, repeated failure, VIP accounts, legal risk.
- QA sampling beats blindly chasing containment rate.
- Embed AI as an agent assistant before full replacement fantasies.
Can doWith sources
Must escalateWrite into policy
GovernanceRefuse over invent
| Decision lens | Common trap or cost | Safer cut |
|---|---|---|
| Can do | FAQ/routing | With sources |
| Must escalate | Promises/refunds/emotion | Write into policy |
| Governance | QA sampling | Refuse over invent |
Policy boundaries before model flexing
- Promise AI replaces agents?
- Do not promise customers. Cover nights and repetitive FAQs first.
- Multilingual?
- Maintain language-specific knowledge; human-sample critical categories.
- Ticket system hookup?
- Auto-create tickets, write summaries back, tag for routing.
- Control hallucinations?
- RAG, refusal policies, ban inventing policy clauses.