Help centers fail when search returns ten articles and users pick the wrong one. Cited AI flips the workflow: users ask in natural language, get a direct answer, and see which article or PDF page it came from. Support leads get fewer “where did this answer come from?” escalations because sources are visible by design.
When this workflow fits
Strong fit if you have:
- A mature help center (Intercom, Zendesk, Notion public pages, custom CMS)
- Repeatable questions about setup, billing, or policies
- Docs that change often enough that static FAQs fall behind
Weak fit if answers require live account data — pair cited doc answers with your existing support tools for account-specific issues.
Step 1 — Export or sync your help content
Start with the articles that drive ticket volume. Export to PDF or Markdown, or upload the canonical versions your team maintains (not stale Zendesk duplicates).
Organize uploads by product area so you can tune projects later — e.g., “Billing FAQ” vs “Integrations”.
Step 2 — Validate with support-owned questions
Use the dashboard to run questions pulled from last month’s tickets. Support managers should sign off on:
- Answers that are safe to show customers without editing
- Cases where the bot should defer to a human (account locks, refunds above threshold)
See validate document Q&A before launch for the full checklist.
Step 3 — Choose embed vs custom UI
Two paths:
- Embed widget — Fastest path; drop a script tag on help center pages. See embed a cited chat widget.
- Custom UI + REST API — Match your design system; you control layout of answers and citations. See REST integration guide.
Both run on Oprag’s AWS with tenant isolation per project — your help content is not commingled with other customers.
Step 4 — Place it where deflection actually happens
Put the assistant on:
- Top of help center home
- Article sidebar for related questions
- Post-checkout “Need help?” panels
Avoid burying it on a standalone page nobody visits. Measure deflection by comparing ticket creation rate on pages with vs without the widget.
Step 5 — Train agents on citations
Agents should treat citations as first-class UI:
- Click through to confirm before pasting into replies
- Flag wrong citations as doc bugs, not “AI issues”
- Escalate when citations are missing — often a sign the article does not exist yet
Metrics that matter
Track weekly:
- Self-serve rate — questions answered without ticket creation
- Citation click rate — proxy for trust
- Top unanswered questions — your content backlog
Rollout pattern
- Internal-only on staging help center
- 10% of traffic on high-volume FAQ pages
- Full rollout after golden question set passes
For deeper ticket reduction tactics, read reduce support tickets with cited AI.