The three layers of AI in customer support
- Deflection — chat / email AI that fully resolves tickets without a human
- Agent-assist — AI sitting next to human agents, drafting responses + surfacing relevant docs
- Voice intake — AI answering inbound calls, qualifying, routing
Mature programs run all three. Most teams start with #1 or #2.
Ticket deflection
Best-in-class platforms (Intercom Fin, Zendesk AI, custom Claude builds) handle 40–65% of incoming tickets end-to-end in 2026. Key variables that move that number:
- Quality of training data (your real KB beats generic FAQs by 30+ points)
- Confidence-thresholding (better to escalate gracefully than respond wrongly)
- Multi-turn capability (one-shot Q&A caps at ~25% deflection)
- Tool access (looking up order status, processing refunds, etc.)
Agent-assist
The under-rated win. AI sits next to your support agents, drafts responses, surfaces relevant macros, summarizes long ticket threads. Typical impact:
- 30–45% reduction in handle time
- 15–20% improvement in CSAT (agents have more time to be human)
- 20–35% faster onboarding for new agents
- Tooling: Cresta, Forethought, Ada, custom builds on Salesforce / Zendesk
AI voice for support
Voice support is the fastest-growing AI deployment in 2026. Best use cases:
- After-hours coverage
- Tier-1 status questions ("Where's my order?")
- Appointment changes
- Account-state lookups
- Smart routing to the right human team
Full guide: AI Voice Agent Guide 2026.
Metrics that matter
| Metric | Target |
|---|---|
| Deflection rate | 40–65% (mature programs) |
| CSAT on AI-only conversations | ≥ baseline human CSAT |
| Escalation rate (false negatives) | < 15% |
| Cost per resolved ticket | $0.30–$1.10 (AI) vs $4–$15 (human) |
| Time-to-first-response | < 30 seconds |
Want a customer-service AI plan tailored to your stack? Book a call.