Turn the tickets you keep answering into a knowledge base that answers them for you

Your support queue has a memory problem, and it isn't your team's. The same forty questions arrive every month — the expired-session logout, the empty export, the SSO setup that never quite lands — and every one gets a thoughtful, correct answer that then vanishes into a closed ticket no customer will ever read. The knowledge exists. It's just trapped in resolution threads, one reply at a time, waiting for someone with a free afternoon to dig it out and write it up. That afternoon never comes, because that someone is answering question forty-one.

This workflow gives that afternoon to a schedule instead. Every Monday at six, an internal:schedule trigger wakes the run, and a Zendesk node pulls the week's resolved tickets — subject, tags, and the resolution thread where the real answer lives. A filter node throws out everything that can't become an article: the one-touch closes, the spam, the tickets already covered. What's left is pure signal — real problems with real answers already attached.

Then the clever part. A Milvus node embeds every surviving ticket and does two things at once. It dedupes — a ticket that's already covered by a published article gets suppressed, so you never redraft what exists. And it clusters — forty scattered password-reset tickets collapse into one high-signal group. The workflow stops seeing forty lonely rows and starts seeing what your team feels but can't quite measure: this is a recurring problem, and it's forty tickets deep.

Each cluster goes to an AI agent node running Claude, which reads the representative resolutions and writes the article your team never had time to — a plain-language title, the symptom, the steps, an escalation note, all written for a customer who has never seen the internal thread. A structured parser pins that draft to a clean schema so the next node gets fields, not a wall of prose. Then a Notion node creates the page in your knowledge-base database, marked Draft, Needs Review. Nothing customer-facing ever ships on its own.

A Slack node closes the loop with a card in your review channel: here's the article, here's that it came from forty tickets this week, here's the model's confidence, here's the link. Your support lead reads it, tweaks a line, hits publish — the one deliberate human act in the whole run. And here's the quiet magic: next Monday, Milvus recognizes that new published article and suppresses its duplicate tickets. The loop tightens itself. The knowledge base grows from the exact questions your customers keep asking, and the questions it answers stop coming back.