Same 12 issues drive 60% of support tickets. KB articles would deflect them — but writing KB articles takes time, so they don't exist, so the tickets keep coming.
The AI approach
Cluster solved tickets by problem; LLM reads the cluster + actual fixes used; drafts a Problem / Cause / Fix / FAQ article. Support reviews and publishes.
The outcome
~−18% tickets
Top 12 articles deflect ~18% of tickets within 30 days. Articles based on real solved tickets, so they stay accurate.
Try itInput → Process → Output
Input — ticket cluster (solved)
Cluster: "Promo code not applying"2,103 solved tickets · last 90 days
"discount field never works" "voucher won't apply at checkout" "3rd time my SAVE10 code fails" "why does the promo never work??"
Resolution notes · most common fix
91% resolved by: "check that the cart subtotal exceeds the promo's minimum (most are ฿500)" · 6% via "code expired" · 3% genuine bug
Process — AI pipeline
1Cluster solved tickets by problemReadPredictive
2Read cluster + resolutionsReadGenerative
3Draft KB article (P/C/F/FAQ)WriteGenerative
4Apply KB style + link checkRulesSymbolic
Output — KB article draft
Click Run demo to draft a KB article from the cluster.
Your promo code (e.g. SAVE10) doesn't apply at checkout, even though the code is valid.
Cause (most common)
Most promo codes have a minimum order value — usually ฿500 — that has to be met before the discount is applied. The error message doesn't always make this clear.
Fix
Check the subtotal at the top of your cart.
If it's below the promo's minimum, add items until you cross it.
Re-enter the code at checkout.
Cluster basis: 2,103 tickets · 91% match this fix path. Article scheduled to be reviewed by support team before publishing to Help Center.
Three AI types in this use case
SymbolicKB style template (Problem / Cause / Fix / FAQ); link-validity checker; obsolete-content flag (when a fix changes upstream).
PredictiveEmbedding + clustering (HDBSCAN) on solved tickets; cluster size + recency for prioritisation.
GenerativeLLM reads ticket cluster + resolution notes; writes the article in standard KB voice.
The stack
Helpdesk export · solved tickets only
Embeddings · cluster by problem
LLM · KB voice prompt
KB · Zendesk / Notion
When this works
Tickets are well-tagged or text-rich enough to cluster
Resolution notes captured
Support reviews + updates
When it fails
Two issues with same symptom but different fix get merged