Context
A B2B SaaS at $15M ARR. 12-person support team across two time zones. Pre-AI state: 5,000 tickets/month, average resolution time 16 hours, CSAT 78%. New tickets growing 30% YoY ahead of revenue.
The Audit
500-ticket classification revealed:
- 42% — simple status / how-to (could be auto-resolved).
- 31% — needs human but AI could draft the reply.
- 18% — complex, needed senior agent.
- 9% — not actually support tickets (sales, partnerships, billing edge cases).
The Build
Eight-week phased rollout:
- Build a grounded knowledge base from product docs and past tickets.
- Deploy AI triage classifier (routes to one of five paths).
- Open auto-resolve for top 10 ticket types with aggressive escalation.
- Launch AI drafting for the human-handled tier.
- Add senior side-tools (summarization, draft-the-difficult-paragraph).
Results
- Support cost per ticket: -47%.
- Average resolution time: 16h → 4h.
- CSAT: 78% → 84%.
- Team size: shrunk by 4 (through attrition, not layoffs).
- Senior agent escalation rate: stable (no quality erosion).
- • 47% support cost reduction.
- • Higher CSAT (faster resolution beats human-everywhere).
- • Senior agents now focus on complex/strategic tickets.
- • Junior agent burnout vanished.
Why CSAT Went Up
Customers preferred fast accurate AI replies for simple questions over slow human ones. The complex tickets got better human attention because senior agents had time. The disclosed AI didn't bother customers; the slow response time used to.
What Broke
- The first knowledge base was too broad; the AI hallucinated. Rebuilt with stricter canonicalization.
- Auto-resolve was too aggressive on edge cases for one product line. Tightened thresholds.
- One enterprise customer escalated about “AI handling our tickets.” Built an opt-out for enterprise tier.
Lessons
- Knowledge base quality determines everything. Invest before the AI.
- Disclosure and easy escalation are non-negotiable.
- Senior agents need to be coached into their new role.
- Enterprise customers may need an opt-out lane.
The 47% cost reduction wasn't the headline. The headline was a calmer support team producing better outcomes for customers. The economics followed.
FAQ
Cost? ~$80k build + $4k/month run cost. ROI in quarter two.
What about churn risk from AI handling support? Did not materialize when disclosure was clear and escalation easy.
Did the team feel threatened? Yes at first. CEO framing as “you handle the harder, more interesting work” calmed it.