Workflow: auto-create KB articles from support tickets

Recommendation
Build a Zapier → ChatGPT (OpenAI) pipeline that: 1) triggers on resolved Zendesk tickets, 2) generates a concise KB draft + metadata, 3) checks for duplicates using a semantic match (lightweight or full vector search), and 4) creates a draft in Zendesk Guide and notifies an editor for quick review. Keep automated drafts limited to low-risk categories (how-tos, configuration steps) and always require a human sign-off before publishing.

Why this works
You get repeatable automation that reduces writer time without risking incorrect content being published. Zapier provides the glue for events/notifications; ChatGPT (OpenAI) handles drafting and summarization. The main trade-off: higher-quality duplicate detection needs a vector store/embeddings and modest infra; simpler fuzzy/title checks are cheaper but noisier.

Decision criteria (pick your path)
- Budget & scale: If budget is tight, start with title/keyword + fuzzy text matching in Zapier (Formatter, Google Sheets lookup). If you have budget for accuracy at 2k tickets/mo, invest in embeddings + vector DB (Pinecone/Weaviate/Vectara) to do semantic duplicate checks.
- Skill level: Low-code teams stick to Zapier built-ins + Zendesk Guide app. Dev teams can add an embedding step and a small service to host similarity queries.
- Output quality: If articles are customer-facing legal/financial content, require full human rewrite. For procedural how‑tos, a lightly edited AI draft is fine.
- Team size & workflow stage: Small teams rely on a single editor notification; larger teams should route by topic tags and set SLAs.

Practical checklist (Zap structure)
1) Trigger: Zendesk ticket status = Solved (or tag = publish_candidate).
2) Pre-filter: Exclude internal-only tickets, private notes, or low-value categories.
3) Create a concise extraction: call ChatGPT to produce: title suggestion, 1-paragraph summary, step-by-step solution, tags, suggested keywords, related tickets keywords. Use a strict prompt (see below).
4) Duplicate check:
- Lightweight: compare title/keywords against Help Center search or a maintained Google Sheet of titles (Zapier lookup + fuzzy match).
- Robust: create an embedding of ticket text (OpenAI embeddings), query vector DB for similarity threshold (e.g., cosine > 0.78) to find near-duplicates.
5) If duplicate found: add a note to ticket, suggest linking to existing KB, and stop. If not, create a draft article in Zendesk Guide via Zapier.
6) Notify: send Slack/Email to editor with draft, highlights, confidence score, and a one-click link to publish.
7) Logging: append metadata to an audit Google Sheet or Airtable (ticket id, draft id, similarity score, editor assigned).

Prompt blueprint for ChatGPT (brief)
"Write a KB article from this Zendesk ticket. Output: suggested title (short), 1-sentence summary, 3–7 step numbered solution, 2–3 tags, one-sentence TL;DR for support agents. Keep language plain, limit to ~250–400 words. Ticket: {ticket_subject}
Conversation: {ticket_comments}"

Best-for / Avoid-if
- Best for: repetitive how-to fixes, setup guides, common configuration questions, FAQs. Works well if you want drafts to speed editors.
- Avoid if: one-off custom dev requests, legal/medical content, or anything needing precise policy language without full human review.

Final tip
Start with the simple Zapier fuzzy-match flow, iterate by adding embeddings and a vector DB once you see patterns. Use the audit log to tune similarity thresholds and reduce false positives.

Compare Zapier and Make