Recommendation (short):
Use Zapier to trigger on new support email, run a deterministic PII-redaction + rule-based classifier, call ChatGPT for structured summary + tag/priority suggestions via a strict JSON prompt, validate the JSON in Zapier, and route low-confidence cases to a human review queue in Zendesk. Start conservative (redact, human-review on low confidence) and iterate.

Proposed Zap flow (step-by-step):
1) Trigger: New email (Gmail/IMAP) or New Zendesk ticket.
2) Formatter/Code step: PII detection & redact. Apply regex for emails, phones, credit cards, SSNs; replace with tokens like [EMAIL_1]. Store a secure mapping (ID -> original) in an internal DB or private Zendesk note.
3) Rule-based pre-check: run fast keyword rules for high-certainty labels ("refund", "cancel", "chargeback", account numbers). If rule matches, set a rule_tag and priority override.
4) Webhook to ChatGPT (OpenAI API via Webhooks by Zapier or the OpenAI app): send a concise prompt asking for a strictly-formatted JSON response (see sample below). Include a few short examples (few-shot) to reduce misclassification.
5) JSON validation: use Zapier’s “Code by Zapier” or Formatter to validate output against a JSON schema. If parsing fails, send to human review.
6) Confidence branching: if model returns confidence below threshold (or model output conflicts with rule_tag), create a Zendesk ticket in “requires-triage” group instead of auto-assigning.
7) Final action: update/create Zendesk ticket with summary, tags, priority, and private note linking to redaction mapping.
8) Logging & retries: log raw responses (masked) and implement retry/backoff for webhook failures.

Prompt + output controls (practical):
- Ask for EXACT JSON: {"priority":"low|medium|high|urgent","tags":["billing"],"summary":"one-sentence","confidence":0-1}
- Provide 3 short labeled examples.
- Ask model to refuse to output any detected PII and note which fields were redacted.

Decision criteria (pick based on budget/needs):
- Accuracy vs cost: use GPT-4 for higher accuracy if budget allows; gpt-3.5 for cheap/high-throughput.
- Team size/workflow stage: small teams should be conservative (more human review). Large teams can auto-assign with higher confidence thresholds.
- Output quality: require JSON schema validation and examples if you need predictable tags for routing.

Best-for / Avoid-if
- Best for: teams who want fast, consistent triage and summaries but keep humans for edge cases.
- Avoid if: you must never expose full PII to any third-party model — redact fully and keep only masked context.

Practical checklist before launch:
- [ ] Implement robust regex + ML-based PII redaction and mapping storage.
- [ ] Build keyword overrides for critical cases.
- [ ] Craft few-shot prompt and strict JSON schema; test 200 historical emails.
- [ ] Add JSON validation and fallback-to-human path.
- [ ] Monitor misclassification rate weekly; iterate prompts and rules.
- [ ] Choose model by budget (gpt-4 for accuracy; gpt-3.5 for scale).

If you want, I can give a ready-to-drop prompt + JSON schema and a Zapier step mapping for Webhooks -> Zendesk.

Compare Zapier and Make