Workflow Blueprint Template

Use this template to design AI-native workflows that are auditable and resilient. Output: 1 diagram + 2-page blueprint.

What Is an AI-Native Workflow?

An AI-native workflow embeds AI into the core decision path, not "off to the side." It includes:

1. AI task (what the system decides)
2. Human checkpoints (where humans verify or override)
3. Stop cards (when humans must intervene)
4. Handoff rules (when to escalate or defer)
5. Audit trail (proof of decision integrity)

Step 1: Workflow Diagram (High-Level)

┌─────────────────────────────────────────────────────────────────┐
│                     SUPPORT TICKET INTAKE                       │
└─────────────────────────────────────────────────────────────────┘
                              ↓
         ┌────────────────────────────────────────┐
         │   AI: Classify Ticket Category          │
         │   (Prompting Skill)                     │
         │   ✓ Input: Ticket text                  │
         │   ✓ Output: Category + confidence       │
         └────────────────────────────────────────┘
                              ↓
         ┌─────────────────────────────────────────────────────┐
         │  Decision Logic (Critical Evaluation Skill)         │
         │  IF confidence > 0.85:                              │
         │    → Route automatically (non-urgent only)          │
         │  IF confidence 0.65-0.85:                           │
         │    → Flag for human review                          │
         │  IF confidence < 0.65:                              │
         │    → STOP CARD: Human review required              │
         │  IF category = "Safety/Fraud":                      │
         │    → STOP CARD: Escalate immediately               │
         └─────────────────────────────────────────────────────┘
                              ↓
    ┌────────────────────┬──────────────────┬──────────────────┐
    │                    │                  │                  │
    ▼                    ▼                  ▼                  ▼
┌─────────────┐    ┌─────────────┐    ┌─────────────┐    ┌──────────────┐
│ Auto-Route  │    │ Queue for   │    │ Human      │    │ Escalate to  │
│ (Confidence │    │ Review      │    │ Review     │    │ Manager      │
│ >0.85)      │    │ (0.65-0.85) │    │ (<0.65)    │    │ (Safety/     │
│             │    │             │    │            │    │ Fraud)       │
└─────────────┘    └─────────────┘    └─────────────┘    └──────────────┘
         ↓                  ↓                ↓                   ↓
    ┌────────────────────────────────────────────────────────────────┐
    │  Audit Trail: Log Decision, Outcome, Override (if any)        │
    │  (Workflow Integration & Ethics & Trust Skills)                │
    └────────────────────────────────────────────────────────────────┘

Step 2: Detailed Workflow Blueprint

1. Workflow Name & Purpose

Name: Customer Support Ticket Triage
Owner: [Name/Team]
Version: 1.0
Last Updated: [Date]
Goal: Route 80% of tickets automatically, maintain <2% misdirection rate

2. Inputs & Outputs

3. AI Task Definition

What the AI does:

Classify a customer support ticket into 1 of 8 predefined categories:
1. Billing & Payment
2. Technical Support
3. Account Access
4. Product Feedback
5. Complaint / Escalation
6. General Inquiry
7. Other
8. Safety / Fraud Alert

Model: Claude 3.5 Sonnet (via Anthropic API)
Latency SLA: <2 seconds per ticket
Cost: ~$0.001 per classification

What the AI does NOT do:

• Draft responses to customers (humans do this)
• Decide ticket priority or SLA
• Close tickets without human sign-off
• Contact customers directly

4. Decision Logic & Stop Cards

4a. Confidence-Based Routing

**IF confidence_score > 0.85 AND category != "Safety/Fraud":**
  └─ AUTO-ROUTE to appropriate queue
     └─ Update ticket status: "Routed (automated)"
     └─ Log decision with confidence score
     └─ Human reviewer spot-checks 5% of auto-routed tickets daily

**ELIF 0.65 < confidence_score <= 0.85:**
  └─ FLAG FOR HUMAN REVIEW (add to review queue)
     └─ Show model's top 2 category suggestions
     └─ Show confidence score
     └─ Wait for human decision (<4-hour SLA)

**ELIF confidence_score <= 0.65:**
  └─ STOP CARD: MANUAL REVIEW REQUIRED
     └─ Route to most experienced agent
     └─ Flag as "uncertain—human expertise needed"
     └─ SLA: <2 hours

**IF category == "Safety/Fraud":**
  └─ STOP CARD: ESCALATE IMMEDIATELY
     └─ Notify manager + compliance team
     └─ Do NOT auto-route
     └─ SLA: <15 minutes

4b. Stop Card Conditions (Non-Negotiable Escalation)

5. Handoff Points (Workflow Integration Skill)

5a. AI → Human Review

Trigger: Confidence 0.65–0.85
Responsibility: Human support agent
SLA: <4 hours
Output: Corrected category + reason for correction (logged)

5b. Human Review → AI Learning Loop

If human decision differs from model:
  1. Log disagreement (category, confidence, correction)
  2. Weekly: Aggregate corrections
  3. Monthly: Retrain prompt or fine-tune model if >5% pattern
  4. Update prompt library

5c. Escalation → Manager

Trigger: Stop card conditions (safety, model drift, system failure)
Responsibility: Support manager
SLA: Varies (safety = 15 min, others = 1 day)
Output: Incident report + corrective action

6. Audit Trail & Evidence (Ethics & Trust Skill)

What gets logged for every ticket:

{
  "ticket_id": "TKT-12345",
  "timestamp": "2026-02-06T14:23:45Z",
  "input": {
    "subject": "Can't reset my password",
    "body": "I've tried three times. Says 'invalid token.' Please help.",
    "customer_account_type": "premium"
  },
  "ai_output": {
    "category": "Account Access",
    "confidence": 0.92,
    "reasoning": "Customer explicitly mentions password reset failure."
  },
  "routing_decision": {
    "auto_routed": true,
    "queue": "account-support",
    "confidence_threshold_met": true
  },
  "human_override": {
    "occurred": false,
    "override_reason": null,
    "human_category": null
  },
  "outcome": {
    "actual_category": "Account Access",
    "resolved_correctly": true,
    "resolution_time_minutes": 12
  }
}

Retention: 2 years minimum (for regulatory audit)
Access: Logged-in support agents + compliance team
Security: Encrypted at rest, masked PII in logs

7. Monitoring & KPIs (Critical Evaluation Skill)

Daily Dashboard

Weekly Review

• Sample 100 auto-routed tickets; verify category accuracy
• Aggregate human overrides; identify patterns
• Check for category drift (e.g., more "Safety" alerts than before?)
• Review stop card triggers; were they valid?

Monthly Review

• Retrain/refine prompt if accuracy <93%
• Benchmark against baseline (human-only routing)
• Calculate cost savings: (4 FTEs saved) - (model cost + human review overhead)
• Report to business stakeholder + compliance team

Step 3: Integration Checklist (AI Strategy Skill)

• Workflow goal is measurable (e.g., "reduce ticket routing time from 8 min to <1 min")
• AI task is specific & scoped (classify ≠ decide priority ≠ respond)
• Confidence thresholds are defined (>0.85 auto, 0.65–0.85 review, <0.65 escalate)
• Stop cards are crisp & unambiguous (safety, drift, failure = immediate escalation)
• Handoff rules are clear (AI → human review: <4 hrs; human review → AI learning: monthly check)
• Audit trail captures everything (input, output, override, outcome, timestamp)
• Monitoring is automated (daily dashboard, weekly spot-check, monthly full review)
• Escalation path is documented (who gets paged for safety? for drift? for outage?)

Step 4: Pilot & Validation

Pilot Phase (Week 1)

• Deploy to 5% of traffic (100 tickets/day)
• Monitor accuracy daily
• Document all human overrides + reason
• Alert if any Stop Card triggered

Go/No-Go Decision:

• If accuracy ≥92%: proceed to expansion
• If accuracy <92%: debug prompt, retest

Expansion Phase (Weeks 2–4)

• 25% → 50% → 100% of traffic (if pilot successful)
• Monitor key metrics daily
• Implement weekly review process

Production Phase (Ongoing)

• Continuous monitoring per KPIs
• Monthly prompt refresh
• Quarterly business review with stakeholders

Foundational Skills Checklist

• AI Strategy: Workflow goal tied to business metric (e.g., cost savings, faster resolution)
• Prompting: Prompt is versioned and tested for accuracy across models
• Workflow Integration: Handoff rules, stop cards, and audit trail fully designed
• Critical Evaluation: KPIs defined, monitoring dashboard built, spot-check process documented
• Ethics & Trust: Guardrails (confidence thresholds, stop cards) designed to survive model changes; audit trail captures all decisions

Real-World Example: Comparison

Workflow Anti-Pattern (❌ Not AI-Native)

Tickets come in → AI classifier runs in background → 
Sometimes teams see the result, sometimes they don't → 
No one knows if accuracy is good → 
AI quietly misdirects critical issues → 
Customers complain

Workflow Best Practice (✓ AI-Native)

Tickets come in → AI classifier + confidence score → 
IF high confidence → auto-route + log → weekly spot-check
IF low confidence → human review (with AI suggestion) → 
IF safety flag → escalate immediately → audit trail captured
Monitoring dashboard shows accuracy, override rate, SLA metrics
Monthly: Review outcomes, retrain if needed, celebrate wins

The difference: Intentional human oversight + continuous learning + measurable outcomes.