Customer Support Automation

Architecture for production support automation: **ingest** (incoming channels — email, chat, ticketing, voice transcripts) → **normalize** (one schema across channels) → **enrich** (user profile, past tickets, purchase history, entitlements) → **classify** (urgency, category, sentiment, language) → **route** (deflection vs human queue) → **respond** (RAG for deflection, pre-draft for copilot, escalation with summary for handoff).

**Knowledge base quality is the ceiling**. Your RAG deflection accuracy is bounded by the KB — bad docs = bad answers, missing articles = hallucinations. Investment sequence: (i) audit KB for currency + coverage; (ii) add 'article request' workflow from agents when the bot fails; (iii) close the loop — every escalation becomes a KB candidate.

**Classifier patterns**: small fast model (Haiku / GPT-nano / fine-tuned BERT) for classification; frontier model (Sonnet / GPT / Gemini Pro) for generation. The classifier outputs a structured decision; you gate further LLM calls on the classification (skip generation for 'needs_human' class).

Escalation + handoff quality is the real differentiator. When the AI escalates, it should pass: (a) summary of what was tried, (b) what the user asked, (c) relevant KB articles already referenced, (d) emotional state / urgency flags. Agent sees a pre-contextualized ticket, saves 2-5 min of reading backstory.

Guardrails: refusal rules for (i) actions with financial impact (refunds, credits — require agent approval); (ii) account-modifying actions (plan change, cancellation — require authenticated confirmation); (iii) sensitive topics (legal threats, safety, self-harm — immediate human escalation); (iv) off-topic (competitor comparison, product roadmap — templated response). Without guardrails, the bot will get creative and create problems.

**Evaluation infrastructure**: sample 5-10% of AI-handled tickets weekly, grade (correctness, tone, completeness) via LLM-as-judge + occasional human audit. Track CSAT delta (AI-handled vs human-handled), deflection rate, false-positive escalation (the AI escalated when it could have handled), false-negative escalation (the AI tried when it should have escalated — the dangerous one).

**Voice channel specifics**: real-time transcription (Whisper, Deepgram, AssemblyAI) + low-latency turn-taking. Frontier voice models (OpenAI Realtime, Gemini Live, Anthropic voice) handle barge-in, tone, interruptions natively. Latency budget: < 800ms round-trip for natural conversation.

Integration surface: ticketing systems (Zendesk, Intercom, Freshdesk, Salesforce Service Cloud, Gorgias, Kustomer), CRMs, KBs (Guru, Notion, Confluence), product internal APIs (for account context), BI / observability (all actions logged for analysis).

Common anti-patterns: (i) deflection-only strategy without human fallback — tanks CSAT; (ii) replacing human agents with AI without retraining — attrition + bad AI because no one to improve the system; (iii) deploying without guardrails — one refund loophole = viral Twitter moment; (iv) measuring wrong metrics (volume handled, not satisfaction).