Every new Zendesk ticket gets a category, priority and sentiment tag plus a draft reply in an internal note before an agent opens it.
Agents spend the first few minutes of every ticket working out what it is about, who should own it and where the answer lives. Urgent issues like failed payments or outages sit in the same queue as feature requests until someone reads them. First reply times slip every time volume spikes after a release.
In Zendesk Admin Center, create a webhook that points at an n8n Webhook node, then a trigger on "Ticket is Created" that notifies it with ticket ID, subject, description, requester email and organization.
Write down 8–12 categories that match your existing Zendesk groups (billing, login, bug, how-to, feature request, cancellation) and your priority rules. For example, any mention of payment failure, data loss or the product being down for several users is urgent.
Add a Basic LLM Chain with an Anthropic Chat Model node (Claude Haiku), the triage prompt and a Structured Output Parser for category, priority, sentiment, language and confidence. Classification does not need a larger model, which keeps cost per ticket at fractions of a cent.
Use an HTTP Request node to query the Zendesk Help Center search API with the ticket subject and keep the top three articles. Pass their titles, URLs and body text to the drafting prompt along with any macro mapped to the category.
Update the ticket with tags such as ai_cat_billing and ai_pri_urgent, set group and priority, and add the draft as an internal note (public: false). Agents edit and send it themselves, so nothing reaches a customer unreviewed.
Build a Zendesk view of tickets where an agent changed the AI category and review it weekly. When confidence comes back below 0.6, skip the draft and tag the ticket ai_needs_human instead.
You triage support tickets for {{company_name}}, which makes {{product_description}}.
Categories: {{category_list}}
Urgent means: payment failures, security concerns, data loss, or the product being unusable for more than one user.
Ticket subject: {{subject}}
Ticket body: {{description}}
Customer plan: {{plan}}
Return only JSON:
{
"category": one of the categories above,
"priority": "urgent" | "high" | "normal" | "low",
"sentiment": "angry" | "frustrated" | "neutral" | "positive",
"language": ISO 639-1 code,
"confidence": number from 0 to 1,
"summary": "one sentence for the agent"
}
If the ticket covers two issues, classify the one that blocks the customer more.Write a first reply to this support ticket for an agent to review and send.
Ticket: {{description}}
Customer name: {{requester_name}}
Category: {{category}}
Sentiment: {{sentiment}}
Relevant help articles: {{articles}}
Matching macro, if any: {{macro}}
Rules:
- Use only information from the articles or macro. If they do not cover the question, say what you will check and do not invent steps or timelines.
- Link at most two articles, using their exact URLs.
- Reply in the customer's language ({{language}}).
- Under 150 words, plain sentences, no filler openers.
- If sentiment is angry or frustrated, acknowledge the specific problem in the first sentence.
End with a line starting "AGENT NOTE:" listing anything the agent must verify before sending.Score every demo request against your ICP, update HubSpot and ping the right rep in Slack within a minute of the form submit.
Spot order-status questions in the support inbox, look up the order in Shopify and reply with live tracking and an honest ETA.
Each month, find posts whose search clicks dropped, research what changed with Perplexity, and get a Claude-written refresh brief for each one.