Intercom Fin AI for support deflection (L3)
Deploy Intercom Fin (or Ada/Zendesk AI) trained on your docs + past ticket history. Fin answers ~50\,70% of support tickets without a human, routes the rest with context attached. Each deflection pays for itself in 2 tickets.
The steps
- 01
Audit and clean your docs first
Tool: Notion
Fin is only as good as its knowledge base. Before connecting it, audit your help center: kill outdated articles, merge duplicates, add a "last verified" date. Aim for <300 high-quality articles, not 2000 stale ones. Owner: Support lead + tech writer. Time: 2,4 weeks. Pitfall: enabling AI on top of a messy KB, you'll auto-resolve tickets with wrong answers and tank CSAT. DoD: 100% of articles in scope have a last-verified date <90 days.
- 02
Configure resolution thresholds + escalation rules
Tool: Intercom
In Fin: set the confidence threshold (start at 0.8, only auto-resolve when AI is very confident). Define escalation triggers: account ARR >$50k → always human; emotional language detected → always human; billing/cancellation requests → always human. Owner: Support ops. DoD: tested escalation rules with 10 synthetic tickets per category; all routed correctly.
- 03
Hand off to humans with full context
Tool: Intercom
When Fin escalates, the human agent sees: full Fin conversation, customer's account data, the 2 most relevant docs Fin tried, AND the reason Fin escalated. The handoff is the difference between "AI saves us money" and "AI makes customers repeat themselves." Owner: Support ops. Pitfall: dropping the AI transcript before handoff, customers will write "I just told the bot this" in their first message to a human, and CSAT craters. DoD: every escalation has the AI conversation pinned to the top of the human-agent view.
- 04
Measure deflection + CSAT separately
Tool: Intercom
Don't celebrate deflection rate alone. Track: (a) % auto-resolved, (b) CSAT on auto-resolved (must stay within 0.3 points of human-resolved), (c) re-open rate on auto-resolved (should be <10%). If CSAT or re-open rate degrades, raise the confidence threshold. Owner: Support ops + analytics. DoD: weekly dashboard with all 3 metrics; documented threshold-tuning history.
Tools in this playbook
Next playbooks
Unfamiliar terms are defined in the AI and Revenue Dictionary. Related frameworks live in the framework library.