Recommendation:
Build a Zap that: (1) triggers on new/updated Zendesk tickets, (2) optionally batches tickets for cost control, (3) sends ticket text to Gemini with a structured prompt + few-shot examples, and (4) writes the parsed action items back to the ticket (internal note or a custom field). Use Zapier to orchestrate triggers, filters and batching; call Gemini via a Webhooks/Custom AI action or your provider’s API step.

Why this works
- Keeps agents focused: 1–3 action items per ticket, owner/urgency hints.
- Batching reduces API calls/costs and lets you include multiple tickets’ context when needed.

Decision criteria (how to pick realtime vs batching)
- Low volume (200/day): batch hourly or every few hundred tickets; include only summary fields to limit token usage.
- SLA sensitivity: if your SLA requires immediate triage, prefer realtime for high-priority tickets and batch the rest.
- Budget & token cost: batch more often to amortize fixed overhead and to use a combined request if many short tickets.
- Skill level: low/no-code teams should use Zapier native integrations + Webhook step. Engineers can add deduping, embeddings, or summarization microservices.

Prompt templates (practical, copy-paste-ready)
System prompt:
You are a ticket-summarization assistant. For each ticket, output a JSON array named items. Each item must have: action (one short imperative sentence), suggested_owner (role, e.g., "support", "engineering"), and urgency (low/medium/high). Keep each action ≤ 12 words. Do not add commentary.

User prompt (single-ticket example):
Ticket ID: {{ticket_id}}
Subject: {{subject}}
Description: {{description}}
Output required in strict JSON exactly like: {"ticket_id":"...","items":[{"action":"...","suggested_owner":"...","urgency":"..."}]}

Few-shot example (include 1–2 examples inline):
Input: "App crashes when saving draft; user on iOS 17; steps: open app → create draft → save → crash."
Output: {"ticket_id":"1234","items":[{"action":"Reproduce crash saving draft on iOS 17","suggested_owner":"engineering","urgency":"high"},{"action":"Reply to user for crash logs","suggested_owner":"support","urgency":"medium"}]}

Batch prompt (when sending N tickets):
Process the following tickets independently and return a JSON object with keys equal to ticket_ids. For each, return items as above. Do not merge tickets.

Zapier Zap checklist (practical step list)
1) Trigger: Zendesk — New Ticket or Updated Ticket.
2) Filter: skip low-value tickets (e.g., automation tags, spam).
3) Formatter/Prepare: combine subject+description, add ticket_id, priority, tags.
4) Delay or Schedule: (optional) for batching frequency.
5) Action: Webhooks/POST to Gemini endpoint with your prompt and examples (or use Zapier AI if available).
6) Parser: Use a “Code by Zapier” or Formatter step to parse returned JSON.
7) Update Zendesk: Add internal note or populate a custom “AI Action Items” field with the structured result.
8) Logging: Save original + AI output to a datastore or Slack channel for QA.

Best-for / Avoid-if
- Best for: triage teams that need concise next steps, smaller ticket text, clear SLAs.
- Avoid if: tickets require long threaded context, back-and-forth clarifications, or you need legally auditable summaries without human review.

Extra tips
- Include ticket metadata (tags, priority) in prompt to improve owner/urgency selection.
- Start conservative: run AI output to a private field for QA for 1–2 weeks before exposing to agents.

CTA: If you want, I can draft the exact Webhook payload and the example zap step-by-step for Zapier.

Compare Zapier and Make