AI ticket triage and routing n8n template

What this workflow does

1. Zendesk ticket created (Webhook) receives the ticket ID, subject, first message, requester, organization, plan tier and tags from a Zendesk trigger.
2. Prepare classification (Code) skips tickets already tagged ai_triaged, so the workflow’s own update cannot loop, and builds the prompt from your taxonomy.
3. Classify with Claude (HTTP Request) calls the Anthropic Messages API with a strict JSON schema. Every enum in the schema comes from the taxonomy, so the model can only answer with values you defined.
4. Validate and set priority (Code) handles refusals, truncated answers and API errors, rejects any value not on the allowed list, and bumps priority one level for top-tier accounts.
5. Classification OK? (If) sends failures to the exceptions channel.
6. Log triage result (Google Sheets) writes one row per ticket so you can measure agreement with your agents.
7. Update Zendesk ticket adds tags (ai_triaged, ai_cat_*, ai_area_*), fills custom fields and, once you allow it, sets priority.
8. Escalate? and Alert urgent ticket post urgent tickets, or negative tickets from top-tier accounts, to Slack with a link.

Before you import

• A current n8n version (cloud or self-hosted) with a public URL for the webhook.
• A Zendesk admin who can create webhooks, triggers and custom fields, plus an API token.
• An Anthropic API key. The template uses claude-opus-5-5; claude-haiku-5-5 is cheaper for high volume.
• A Google Sheet with a tab named Triage log and a Slack bot that is a member of both channels.

Setup

1. Import the JSON and create the four credentials. Select them on the Zendesk, HTTP Request, Google Sheets and Slack nodes.
2. Activate the workflow and copy the production URL from Zendesk ticket created.
3. In Zendesk, create a webhook to that URL (POST, JSON), then a trigger on ticket created with the condition “Tags contains none of ai_triaged”. Paste the JSON body from the sticky note.
4. Edit CATEGORIES and PRODUCT_AREAS in Prepare classification and the matching ALLOWED lists in Validate and set priority.
5. Replace REPLACE_WITH_CATEGORY_FIELD_ID and the other FIELD_IDS with your numeric Zendesk field IDs, or leave them and use tags only.
6. Replace REPLACE_WITH_SHEET_ID, REPLACE_WITH_SLACK_CHANNEL_ID, REPLACE_WITH_SLACK_EXCEPTIONS_CHANNEL_ID and REPLACE_WITH_ZENDESK_SUBDOMAIN.
7. Add this header row to the Triage log tab: received_at, ticket_id, subject, category, product_area, language, sentiment, suggested_priority, final_priority, confidence, low_confidence, escalate, reason.

Test it

Create three test tickets: a billing question, a “site is down” message and a vague one-liner. Check the execution log, the tags on each ticket and the sheet rows. The vague ticket should carry ai_low_confidence. The outage should land in Slack. Then put an invalid key in the Anthropic credential and confirm the next ticket reaches the exceptions channel.

Make it production-ready

• Keep SUGGEST_ONLY = true for a week and compare the logged category with what agents chose before letting the AI set priority.
• Set an n8n error workflow so a failed Zendesk update pages someone instead of failing quietly.
• The HTTP Request node retries three times. Zendesk rate limits depend on your plan, so watch for 429 responses during ticket spikes.
• Store the taxonomy in a sheet once support leads want to edit it themselves.
• Treat ticket text as untrusted input. The prompt says so, and the validator ignores any value outside your lists.

Read the full guide at AI ticket triage and routing for the taxonomy and testing approach, or talk to Lumeor if you want help adapting it to your helpdesk.