Zendesk/ServiceNow Copilot: Draft Replies, Surface Fixes, and Respect Your Macros — A 30‑Day, Governed Rollout

yaml
policy_version: v1.7
owner: support-operations
approved_by:
  - qa_lead: "K. Turner"
  - legal_counsel: "R. Shah"
regions:
  - us-east-1
  - eu-central-1
queues:
  - name: mobile_app
    brand: appco
    sla: { first_response: 15m, resolution: 24h }
    confidence_thresholds:
      suggest_steps: 0.72
      draft_reply: 0.78
    macro_selection:
      allowed_macros:
        - app_crash_ios
        - app_crash_android
      default_macro: app_crash_ios
      rationale_required: true
    escalation_rules:
      to_tier: L2_mobile
      when:
        - low_confidence: draft_reply < 0.6
        - missing_fields: [ os_version, app_version, device_model ]
    pii_controls:
      redact_fields: [ email, phone, card_last4 ]
      log_retention_days: 90
    ab_holdout: 10%
    approvals:
      auto_send_enabled: false
      supervisor_approval_required: true
    notifications:
      slack_channel: "#queue-mobile-app"
      daily_brief: true
  - name: billing
    brand: appco
    sla: { first_response: 30m, resolution: 48h }
    confidence_thresholds:
      suggest_steps: 0.75
      draft_reply: 0.8
    macro_selection:
      allowed_macros:
        - refund_policy
        - invoice_missing
      default_macro: refund_policy
      rationale_required: true
    escalation_rules:
      to_tier: L2_billing
      when:
        - low_confidence: draft_reply < 0.65
        - prohibited_intents: [ chargeback, legal_threat ]
    pii_controls:
      redact_fields: [ full_name, address, card_last4 ]
      log_retention_days: 180
    ab_holdout: 15%
    approvals:
      auto_send_enabled: false
      supervisor_approval_required: true
    notifications:
      slack_channel: "#queue-billing"
      daily_brief: true
routines:
  suggest_steps:
    grounding_sources: [ zendesk_guide, incident_notes, runbooks ]
    max_sources: 4
    explanation: "Show macro ID and why selected; list missing fields."
  draft_reply:
    tone_pack: brand_appco_v2
    guardrails:
      prohibit_phrases: [ "legal advice", "guarantee" ]
      require_signoff: [ refund_over_100usd ]
rbac:
  agent_groups_read: [ tier1, tier2 ]
  agent_groups_write: [ tier1, tier2 ]
  admin_groups: [ support_ops, qa ]
telemetry:
  record:
    - ticket_id
    - agent_id
    - macro_id
    - confidence_scores
    - sources
    - decision: { used_suggestion: bool, edits_count: int }
  export: daily
  sink: observability-vpc

{
  "title": "Zendesk/ServiceNow Copilot: Draft Replies, Surface Fixes, and Respect Your Macros — A 30‑Day, Governed Rollout",
  "published_date": "2025-11-10",
  "author": {
    "name": "Alex Rivera",
    "role": "Director of AI Experiences",
    "entity": "DeepSpeed AI"
  },
  "core_concept": "AI Copilots and Workflow Assistants",
  "key_takeaways": [
    "Put a copilot inside Zendesk/ServiceNow that drafts replies and suggests troubleshooting steps without breaking your macros or SLAs.",
    "Ship in 30 days: Week 1 knowledge audit and voice tuning; Weeks 2–3 retrieval and prototype; Week 4 analytics and expansion plan.",
    "Governed by design: RBAC, prompt logging, confidence thresholds, agent-in-the-loop, and data residency.",
    "Measured impact: faster first response, lower AHT, higher CSAT, and controlled deflection—with A/B holdouts and daily Slack briefs.",
    "Never trains on your data; all interactions are logged with approvals and confidence scores for audit and QA."
  ],
  "faq": [
    {
      "question": "Will the copilot change our macros or tone?",
      "answer": "No. It starts from your macros and approved content, with brand tone packs curated from high‑CSAT tickets. Agents remain in control and can switch macros or edit drafts."
    },
    {
      "question": "How do we make sure it doesn’t hallucinate steps?",
      "answer": "All suggestions are retrieval‑grounded on approved sources only. Each draft shows sources and a confidence score; low confidence routes to safe macros plus a clarifying question."
    },
    {
      "question": "Can we run this in our VPC?",
      "answer": "Yes. We deploy in your VPC or a dedicated tenant. Data residency and redaction are configurable per region and queue."
    },
    {
      "question": "How do we measure impact without gaming metrics?",
      "answer": "We run an A/B holdout, log confidence and usage, and ship a daily Slack brief with AHT, CSAT, deflection, and SLA risk. Supervisors can trace every suggestion to its sources."
    }
  ],
  "business_impact_evidence": {
    "organization_profile": "Consumer mobile app company with 6M MAU; 120 agents across 3 brands using Zendesk; U.S./EU operations.",
    "before_state": "AHT 11m42s in mobile app queue, 8m31s in billing; CSAT trending down; macro adherence <60%; frequent L2 escalations for missing repro steps.",
    "after_state": "Macro-aware copilot live in two queues; macro adherence >85%; agents adopt suggestions in 68% of tickets; incident spikes handled without overtime.",
    "metrics": [
      "AHT down 21% in mobile app and 13% in billing within 30 days.",
      "3.4-point CSAT lift in mobile app queue.",
      "29% fewer L2 escalations for ‘missing steps’.",
      "41% fewer SLA breaches during Monday spikes."
    ],
    "governance": "Approved by Legal/Security due to prompt logging with redactions, RBAC mapped to Zendesk groups, data residency controls (US/EU), human-in-the-loop review, and a strict policy of never training on client data."
  },
  "summary": "Embed a macro-aware copilot in Zendesk/ServiceNow in 30 days. Drafts replies, surfaces fixes, and respects controls—AHT down, CSAT up, fully governed."
}