Headless Support

Headless Support

Overview

Every support ticket costs $15-50 in human time. Deflecting 40% of tickets with self-serve and AI agents isn't just cost savings — it's better customer experience. Customers want answers in seconds, not hours. The mistake: thinking "headless support" means "no support." It means "answers without waiting." This skill covers AI agent deployment, knowledge base architecture, ticket deflection strategy, and the metrics to prove headless support works.

BYOAI / headless stack: Technical teams often pair Attio (programmable CRM) with Plain (API-first support + native MCP) and connect Jesse / Claude Code instead of vendor AI (Fin, Zendesk AI). Load references/byoai-headless-stack.md for integration map, MCP setup, and when to choose Plain vs Intercom.

Authoritative Foundations

• Intercom — Fin AI Agent and Resolution Bot — Fin AI Agent and Resolution Bot
• Zendesk — AI Agents and Answer Bot — AI Agents and Answer Bot
• Plain — API-first headless support and BYOAI via MCP — Model Context Protocol — tool servers for agent-safe CRM and enrichment access.
• Ada — Conversational AI for Support — Conversational AI for Support
• Forethought — AI-First Customer Support — AI-First Customer Support
• Amazon — Working Backwards (deflection reduces cost and improves CSAT) — Working Backwards (deflection reduces cost and improves CSAT)

When to Use

Trigger phrases: "set up AI support agent", "build self-serve support", "ticket deflection strategy", "headless customer support", "automated support", "Fin AI setup", "knowledge base optimization", "reduce support tickets", "chatbot for support", "automated onboarding support", "self-serve portal", "BYOAI support", "bring your own AI support", "headless CRM stack", "Plain support", "Plain MCP", "API-first support", "embed support in app"

The Deflection Funnel

Customer Has Question
    │
    ▼
[Level 0: Product] — In-app tooltips, empty states, contextual help
    │ ~30% resolved here
    ▼
[Level 1: Knowledge Base] — Search, suggested articles, help center
    │ ~25% resolved here
    ▼
[Level 2: AI Agent / Chatbot] — Conversational, trained on KB + past tickets
    │ ~25% resolved here
    ▼
[Level 3: Human Agent] — Complex, sensitive, escalated issues
    │ ~20% (your goal: push this number DOWN)

Target: 80%+ of inquiries resolved WITHOUT human intervention.

Step-by-Step Process

Phase 1: Knowledge Base Architecture

The foundation of headless support. AI agents are only as good as the content they're trained on.

Article architecture (30-50 articles minimum for effective deflection):

Article quality standards:

1. Title = the exact question customers ask: "How do I connect my CRM?" not "CRM Integration Overview"
2. Answer in first paragraph (no fluff, no "In this article we'll cover...")
3. Screenshots for every step (Loom video for complex workflows)
4. "Next steps" section: related articles, contact support if stuck
5. Updated monthly — stale KB articles erode trust

SEO for your own KB (customers use Google to find you):

• Article titles match search intent: "How to export contacts from [Product]"
• Meta descriptions under 155 characters with the answer
• Internal linking between related articles
• Google indexes your help center — make it findable

Phase 2: AI Agent Deployment

Path A — Vendor AI (Intercom Fin, Zendesk AI): Train on help center, configure persona and escalation in-platform. Best when CS team lives in one UI.

Path B — BYOAI (Plain + MCP): Plain as support infrastructure; your agent (Jesse, Claude Code, Codex) connects via Plain MCP (https://mcp.plain.com/mcp). Agent reads threads and help center, drafts with addGeneratedReply, human approves before replyToThread. Load references/byoai-headless-stack.md and mcp-setup for tool scope and write gates. Best for dev-tool products and Attio-style composable stacks.

Pre-deployment checklist (both paths):

• 30+ help center articles published (minimum for effective AI)
• 100+ past tickets reviewed to identify top deflection opportunities
• AI persona defined: "Friendly, expert, concise. Uses customer's name. Admits when it can't answer."
• Escalation path designed: what triggers handoff to human?
• Test suite: 50 real customer questions run through AI, every answer reviewed

AI agent configuration (Intercom Fin / Zendesk AI as models):

AI PERSONA: [Product] Support Assistant

Voice: Friendly, expert, concise (2-3 sentences max per answer)
Rules:
- Answer from help center articles only (don't hallucinate)
- If unsure: "Great question — let me connect you with a specialist who can help"
- Never: guess, make promises, give legal/security advice, be defensive
- Always: use customer's name, link to relevant article, offer human escalation

Escalation triggers (auto-handoff to human):
- Customer types: "talk to human", "agent", "real person"
- Customer frustration detected: multiple rephrases, ALL CAPS, "this is useless"
- Billing issues (high-stakes, emotional — human handles these)
- Security/privacy questions (never AI — legal risk)
- Enterprise customer + P1 issue (revenue at risk = human)

Testing protocol:

1. Run 50 historical tickets through AI agent
2. Score each response: Correct / Partially Correct / Wrong
3. Fix articles for any "Partially Correct" or "Wrong" responses
4. Re-run until 95%+ correct on test set
5. Launch to 10% of customers (canary), monitor for 1 week
6. Ramp to 50%, then 100%

Phase 3: In-Product Self-Serve

Contextual help (Level 0 — before they search):

Principle: Answer the question BEFORE they ask it. Every place a customer could get stuck, put the answer. This is the highest-ROI deflection — it costs nothing and prevents tickets entirely.

Phase 4: Automated Triage and Routing

For tickets that DO reach human agents, automate the triage:

AUTO-TRIAGE RULES:

IF: keywords "can't login", "password", "2FA", "locked out"
THEN: Priority = P1, Assign to = Auth team, Auto-reply = "Our auth team is on this — expect a response within 15 minutes"

IF: keywords "billing", "invoice", "charge", "refund", "cancel"
THEN: Priority = P1, Assign to = Billing team, Tag = billing

IF: keywords "bug", "not working", "error", "broken", "glitch"
THEN: Tag = bug, Assign to = Tier-2 technical queue

IF: keyword "feature request" OR "wish you had" OR "why don't you"
THEN: Tag = feature-request, Assign to = product-feedback (not support queue)

IF: sender is enterprise-tier customer
THEN: Priority = escalate by 1 level, Assign to = dedicated CSM

Auto-responders that set expectations:

P1 (Critical — system down, can't login):
"Got it. Our team is on this right now. You'll hear back within 15 minutes.
Reference: [ticket #]"

P2 (High — feature broken, workflow blocked):
"Thanks for reporting this. We'll have eyes on it within 2 hours. Track
progress: [link to ticket]"

P3 (Normal — question, configuration help):
"Thanks for reaching out. You'll hear from us within 4 hours. In the
meantime, these articles might help: [3 relevant links]"

Phase 5: Measuring Deflection

Core deflection metrics:

Dashboard to track:

HEADLESS SUPPORT DASHBOARD

| Metric | This Month | Last Month | Trend |
|---|---|---|---|
| Total Inquiries | 1,200 | 1,100 | ↑9% |
| Self-Serve (KB) | 480 (40%) | 375 (34%) | ↑6pp |
| AI Resolved | 360 (30%) | 330 (30%) | — |
| Human Resolved | 360 (30%) | 395 (36%) | ↓6pp |
| AI CSAT | 4.2/5 | 4.1/5 | ↑0.1 |
| Human CSAT | 4.4/5 | 4.4/5 | — |
| Cost Saved (vs full human) | $7,200 | $5,900 | ↑22% |

Output Format

HEADLESS SUPPORT PLAN — [Company]

Current State:
- Monthly tickets: X
- CS team size: X
- Current deflection: X% (KB + AI)
- Current cost/ticket: $X

Target State (6 months):
- Deflection target: X%
- AI CSAT target: 4.X/5
- Human tickets to reduce by: X%
- Cost savings target: $X/month

Implementation Phases:
Phase 1 (Month 1): Knowledge Base Audit
  - Audit existing articles for completeness + accuracy
  - Write 20 new articles based on top ticket types
  - Add contextual help to top 5 confusion points

Phase 2 (Month 2): AI Agent Launch
  - Configure AI persona and escalation rules
  - Test with 50 historical tickets
  - Launch canary (10% of customers)

Phase 3 (Month 3): Optimize
  - Review AI performance data
  - Refine articles based on failed deflections
  - Expand to 100% of customers

Phase 4 (Month 4+): Continuous Improvement
  - Weekly review of AI conversations
  - Monthly KB refresh
  - Quarterly deflection rate target review

Implementation Checklist

• 30+ help center articles published with screenshot + step-by-step
• AI persona documented (voice, escalation rules, prohibited topics)
• 50 historical tickets tested through AI — 95%+ correct
• Escalation triggers configured (frustration, billing, security, enterprise)
• In-product contextual help at top 10 confusion points
• Auto-triage rules for ticket routing by keyword, priority, customer tier
• Auto-responders configured for P1/P2/P3 with specific time expectations
• Deflection dashboard live with weekly review cadence
• CSAT survey collects AI-rating vs human-rating separately

Quality Check

Before delivering, verify:

• Output matches the user's stated request
• Named frameworks or sources are reflected in the recommendation
• The deliverable is specific enough for an agent to execute
• Any assumptions, risks, or dependencies are explicit
• No unsupported claims, invented facts, or private/internal references are included

Common Pitfalls

1. AI agent launched too early. 5 help articles and an AI agent = 70% wrong answers, frustrated customers, and damaged trust. Fix: 30+ articles minimum. Test with 50 real questions before launch.

2. No escape hatch. AI agent that can't escalate to human is a customer experience disaster. Fix: Clear escalation triggers. "Talk to human" must always work. Never trap a customer in a bot loop.

3. AI persona mismatch. A chirpy, emoji-filled support bot for an enterprise security product is tone-deaf. Fix: Match AI persona to brand voice. Professional for enterprise, friendly for SMB.

4. Not measuring CSAT per channel. If AI CSAT is 3.8 and human CSAT is 4.5, you're degrading experience to save money. Fix: Track CSAT separately for AI and human. If AI CSAT drops below 90% of human CSAT, pause and fix.

5. Static knowledge base. Articles written once and never updated become wrong, then dangerous. Fix: Monthly KB review. Owner assigned per category. Every product release triggers KB updates.

6. Deflection as the only goal. "100% deflection = 0 support tickets" sounds great but means you're not hearing from customers. Some tickets are valuable product feedback. Fix: Deflect repetitive questions. Keep product feedback and enterprise escalations human-handled.

Execution Artifacts

• references/framework-notes.md — Named frameworks and reference tables
• templates/output-template.md — Deliverable shell for agent output
• scripts/check-output.py — Lightweight deliverable validator
• references/byoai-headless-stack.md — Attio + Plain + MCP BYOAI stack pattern

Related Skills

• support-tool-stack — Intercom, Zendesk, Front, Help Scout comparison and setup
• cs-playbooks — Onboarding, health scoring, CSQLs, churn intervention
• sla-management — SLA design, escalation paths, priority matrices
• cs-analytics-dashboards — CS metrics, NPS, CSAT, health scoring
• customer-onboarding — Structured onboarding, time-to-value, activation