Draft Zendesk replies from your Elasticsearch ticket history

Build an agent workflow that drafts a suggested first reply on every new Zendesk ticket by searching our Elasticsearch index of past resolved tickets and help center articles, and leaves the draft as an internal note so a human agent still owns the response.

Trigger: incoming webhook from Zendesk. A Zendesk admin will configure an outgoing trigger in Zendesk Admin Center under Apps and integrations, Webhooks and Triggers, that fires on 'Ticket Created' and POSTs the ticket id (and ideally subject, description, requester id, tags, brand id, and ticket form id) to the workflow's webhook URL as JSON. The agent should accept this payload as input.

Step 1. Read the new ticket. Use Zendesk Show Ticket with the ticket id from the webhook payload to fetch the full ticket. Pull out the subject and the first comment (the original customer message). If the comment is HTML, strip it down to plain text for searching. Also keep the ticket id handy for posting the internal note later.

Step 2. Search Elasticsearch for similar past tickets and KB articles. Use Elasticsearch Search Documents (Query Index) against the configured index (I will tell the agent the index name and the field mapping during setup). Combine the subject and the body as the query text. Prefer a hybrid query: a multi_match keyword query across the title/body fields, combined with a kNN or semantic query if the index has a dense_vector or sparse_vector / ELSER field configured. Filter to records flagged as resolved or as published KB articles. Return the top 5 hits with their title, snippet, source link, and any resolution summary field. Use a tight _source list so we only get the fields the agent needs.

Step 3. Judge relevance. Read the top 5 hits and decide which ones are genuinely about the same issue the customer is reporting, not just keyword overlap. If fewer than one or two hits look truly relevant, or if the relevance scores are weak and the content does not match the customer's problem, stop the workflow without posting anything. The whole point is to skip silently when there is no real match, so agents are not trained to ignore noisy suggestions.

Step 4. Draft a suggested reply. Write a short, agent-voice draft that: greets the customer briefly, acknowledges the issue in their words, walks through the resolution steps drawn from the matched past tickets and KB articles, and links the KB articles by URL so the human agent can paste them through. Keep it under roughly 200 words. At the bottom of the draft, list the sources you used as a short 'Suggested sources' block with the title and link of each cited past ticket or article so the agent can quickly verify.

Step 5. Post as an internal note. Use Zendesk Update Ticket on the same ticket id, adding a comment with public set to false so it is an internal note, not a public reply. The note body should start with a one-line header like 'Geni suggested reply (draft, not sent)' so agents instantly recognize it, followed by the drafted reply, followed by the suggested sources block. Do not change the ticket status, priority, assignee, or tags. The human agent stays in control of the actual response.

Guardrails: never post a public reply, only internal notes. Never include made-up steps or links that did not come from the Elasticsearch hits. If the search returns zero hits or only weak matches, do nothing. If the ticket already has an internal note from this workflow (check List Ticket Comments for our header string), skip so we never double-post on retries.