Building Trusted Support Bots
For CS leaders and CX Ops past basic AI literacy. Eight chapters across the support-bot surface: the CS AI landscape (Zendesk CX Trends 2026 84% strategic priority, Klarna up-then-down, Salesforce Agentforce 36% F100); containment vs deflection vs resolution math (McKinsey cost stack $9–14 human vs $1.50–3 autonomous); knowledge grounding (Air Canada Moffatt precedent, NYC MyCity, McKinsey 1-in-12 vs 1-in-200+ with retrieval); the 5 escalation triggers (Gallup CSAT-decline #1 reason, Chevy Tahoe $1, warm-handoff design); conversational design + disclosure (EU AI Act Art. 50 Aug 2 2026, Pew 70% prefer disclosure, DPD haiku, 6 design rules + 3-sentence pattern); voice support AI (Talkdesk 25% inbound, Cognigy 250+ F2000, assist vs autonomous architectures, <900ms latency); quality monitoring (Forrester #3 failure, Salesforce 27-point gap, 5 metrics + shadow QA); and the close — 2-quarter rollout playbook with interactive builder, 4 trust trip-wires, and the through-line: AI for volume, humans for value.
8
Chapters
~40 min
Duration
Intermediate
Level
No
Certification
Course Content
The CS AI landscape
Zendesk CX Trends 2026 (84% strategic priority, 71% deployed). Avg containment 38–52% vs top-decile 70%+. Klarna up-then-down (2024 → 2025). Salesforce, Intercom, Zendesk, Cognigy, Talkdesk landscape. 4 failure cases: Air Canada, DPD, NYC MyCity, Chevy Tahoe $1.
Containment vs deflection · the math
Containment ≠ deflection ≠ resolution — only resolution matters to CSAT. McKinsey cost stack: $9–14 human vs $1.50–3 autonomous. The whiteboard formula. Klarna reframed in money: appropriate containment, not maximum containment.
Knowledge grounding
Moffatt v. Air Canada — "Air Canada is responsible for all information on its website, including from a chatbot." NYC MyCity harm. McKinsey 1-in-12 vs 1-in-200+. The grounding stack: retrieval, citations, confidence thresholds. The stale-KB trap.
Escalation triggers
Gallup #1 CSAT-decline reason: "the bot kept me in a loop." Chevy Tahoe $1 case. The 5 triggers: low confidence, sensitive intent, repeated misunderstanding, regulated topic, user request. The 4-rule warm handoff. Forrester #1 failure reason.
Conversational design + disclosure
EU AI Act Article 50 from Aug 2 2026 — transparency obligation. Pew Research: 70% of consumers prefer disclosure. The three-sentence disclosure pattern. DPD haiku case. The 6 design rules: short, no apologies-as-deflection, named, single-purpose, on-rails, no fake humanness.
Voice support AI
Talkdesk 25% inbound + 18% AHT down. Cognigy 250+ F2000. $4/call autonomous vs $6.50 human. Two architectures: agent-assist vs autonomous voice agent. Latency <900ms + barge-in. Build the architecture-per-intent matrix.
Quality monitoring
Forrester #3 failure reason: no QA. Salesforce State of Service: 27-point CX gap between teams with active monitoring vs without. The 5 metrics: misrouted escalation rate, CSAT on AI-touched, repeat contact, transitions/session, abandonment. Shadow QA at 2% sample.
Making it stick: your CS AI roadmap
2 use cases, 2 quarters, 1 quality bar. Q1 = FAQ deflection grounded + all 5 triggers + shadow QA. Q2 = one transactional intent. 4 trust trip-wires not to cross. Interactive 2-quarter roadmap builder included. The through-line: AI for volume, humans for value.