Finding the help articles you are missing, from your own tickets

Your tickets are a running list of questions your help center does not answer well. This playbook groups resolved tickets by topic, checks each topic against existing articles, and produces a ranked backlog of new and outdated articles, each with a draft built from how agents actually solved it.

01the problem and who owns it

Help centers are written at launch and updated when someone has time. Meanwhile the product changes, customers find new ways to get stuck, and agents solve the same problem by hand dozens of times a month. Nobody has a clear view of which missing articles would prevent the most tickets.

The knowledge manager owns the content but sees tickets only in aggregate reports by category, which are too coarse. The detail sits in agent replies that are never turned into documentation.

02what the AI does, step by step

1. Pull resolved ticketsRecent solved tickets are exported with the customer question, the final agent reply, tags, and any linked articles. Internal notes are included where they explain the fix.
2. Cluster by underlying problemThe model groups tickets by the actual problem, not the category tag, so a password question asked eight different ways becomes one cluster with a count.
3. Match clusters to articlesEach cluster is compared against the help center index. The result is one of three labels: covered, covered but outdated or unclear, or not covered at all.
4. Rank the backlogClusters are ranked by ticket volume, handle time, and whether a self-service answer is even possible. Account-specific problems are excluded, since no article will fix them.
5. Draft from real resolutionsFor top clusters, the model drafts an article or a revision using the agent replies that resolved those tickets, with screenshots marked as needed and unverified steps clearly flagged.
6. Track deflection after publishingOnce published, the cluster's ticket volume is tracked against its baseline, which shows whether the article helped or customers still cannot find it.

03systems it connects to

• Helpdesk. Ticket exports or API access from Zendesk, Freshdesk, Help Scout, or Intercom.
• Help center. The article platform, read for matching and written to as drafts.
• Product release notes. To flag articles likely affected by recent changes.

04human checkpoints

• Backlog prioritization. The knowledge manager reorders the ranked list based on product plans the data cannot see.
• Accuracy review. A subject expert or senior agent verifies each draft step by step in the product before publishing.
• Publication. Drafts are never published automatically.

05what to measure

• Ticket volume per cluster. Before and after the article ships.
• Article views and search exits. Whether customers find the article and stop searching.
• Backlog throughput. Articles shipped per month from the ranked list.
• Coverage. Share of ticket volume in clusters labeled covered.

06risks and guardrails

• Customer data in drafts. Agent replies contain names, emails, and account details. Strip them before clustering and check drafts for leftovers.
• Documenting workarounds. Some clusters are bugs. Route those to engineering instead of writing a permanent article about a temporary workaround.
• Wrong steps. Drafts built from replies may describe old interfaces. Testing in the current product is not optional.

07build vs buy

Some helpdesks surface trending topics and suggest articles. Those work for spotting obvious gaps on a small help center.

A custom pipeline is worth it when ticket volume is high, when categories are too coarse to be useful, or when you want drafts built from your agents' real resolutions and deflection measured cluster by cluster.

08related playbooks

Browse every customer support playbook or the full library.