Ticket deflection fails when chatbots hallucinate or hide their sources. Cited AI targets a narrower goal: answer questions that are already in your docs, show where the answer came from, and route everything else to humans. Support volume drops when users trust self-serve — not when you automate empathy poorly.
Map tickets to document gaps
Export the last 90 days of tickets tagged FAQ, billing, or how-to. Cluster by subject:
- If answer exists in docs but users still ticket → discovery problem (search/embed placement)
- If answer missing from docs → content problem (write article, then index)
- If answer needs account context → keep human path; do not force RAG
Only automate clusters where docs are authoritative.
Design for verification, not magic
Every deflected ticket avoided should come with:
- A concise answer
- One or more citations (article + page)
- A clear “Still need help?” path to create a ticket
Agents benefit too — they paste fewer macro guesses when citations confirm policy wording.
Placement beats model size
Put cited Q&A on:
- Checkout and billing pages (refund windows, plan changes)
- Setup wizards (integration steps)
- Error pages linking to troubleshooting docs
A small model with good retrieval on the right page beats a large model buried in a footer link.
Measure deflection honestly
Define successful deflection as: user viewed an answer with citations and did not open a ticket within 24 hours. Avoid vanity “chat messages sent” metrics.
Track alongside:
- CSAT on pages with embed
- Agent handle time when users arrive with citation links
- Top questions with no citations (content backlog)
Operational loop
Weekly support ↔ docs sync:
- Review unanswered or low-confidence queries
- Assign doc updates
- Re-upload and re-run golden questions
Use validate document Q&A before launch as the quality gate before each content push goes live.
Implementation paths
- Help center embed — add cited AI to help center
- Widget on marketing site — embed cited chat widget
Both integrate with Oprag on our AWS; tenant data stays isolated per project.
Set expectations with leadership
Cited AI reduces documentable volume. It does not eliminate escalations, angry customers, or edge cases. Pitch it as faster self-serve and cleaner agent workflows — not headcount removal on day one.