AI Copilot for Zendesk/ServiceNow: Drafts, Troubleshoots, Macros

yaml
version: 1.3
owners:
  product_area: "Support Operations"
  tech_contact: "copilot-platform@company.com"
  qa_owner: "support-qa-lead@company.com"
regions:
  primary: "eu-west-1"
  fallback: ["eu-central-1"]
  data_residency_enforced: true
sources:
  - type: "zendesk"
    objects: ["tickets", "macros", "user_fields"]
    groups_allowed: ["Tier1", "Tier2", "QA"]
  - type: "servicenow"
    tables: ["incident", "kb_knowledge"]
  - type: "kb"
    provider: "confluence"
  - type: "release_notes"
    provider: "git"
redaction:
  patterns: ["email", "access_token", "ip_address"]
  on_ingest: true
  on_prompt: true
retrieval:
  vector_db: "pgvector"
  top_k: 6
  freshness_days: 120
  safety_filters: ["PHI", "PCI", "secrets"]
macro_policy:
  respect_macros: true
  macro_match_threshold: 0.78
  suggest_macro_only_if_confidence_below: 0.6
  allow_macro_extension: true
  allowed_macro_ids: [10021, 10044, 10087]
  disallowed_macro_ids: [99901]
  citation_required: true
reply_generation:
  tone_profile: "support-enterprise-friendly"
  hedge_words: ["may", "might", "could"]
  forbidden_phrases: ["guarantee", "promise", "legal"]
  max_tokens: 500
  include_steps: true
  include_sources: true
  confidence_threshold:
    draft_reply: 0.72
    troubleshooting_steps: 0.65
    summary_only: 0.55
controls:
  human_in_the_loop: true
  auto_send: false
  approval_steps:
    - role: "Tier1"
      action: "accept_or_edit"
    - role: "QA"
      action: "sample_review"
  escalation_rules:
    low_confidence_below: 0.5
    escalate_to_group: "Tier2-SSO"
    notify_channels: ["#support-incidents", "#qa"]
observability:
  prompt_logging: true
  log_fields: ["ticket_id", "agent_id", "macro_id", "sources", "confidence", "edits"]
  retention_days: 365
  slo:
    draft_latency_ms_p95: 1200
    acceptance_rate_target: 0.65
    edit_distance_target: 0.25
ab_testing:
  enabled: true
  buckets: ["control", "copilot_v1"]
csat_risk:
  predictive_model: true
  threshold: 0.7
  action_on_high_risk: "require_QA_review"

{
  "title": "AI Copilot for Zendesk/ServiceNow: Drafts, Troubleshoots, Macros",
  "published_date": "2025-11-26",
  "author": {
    "name": "Alex Rivera",
    "role": "Director of AI Experiences",
    "entity": "DeepSpeed AI"
  },
  "core_concept": "AI Copilots and Workflow Assistants",
  "key_takeaways": [
    "Put the copilot where agents work—inside Zendesk or ServiceNow—so replies draft in-line and macros stay the source of truth.",
    "Week 1–4 plan: voice tuning, retrieval setup, prototype in production views, and usage analytics that tie to CSAT/AHT.",
    "Governance is built-in: human-in-the-loop, prompt logs, RBAC, residency, and never training on your data.",
    "Expect measurable impact fast: 20–30% AHT reduction in pilot queues and a 3–5 point CSAT lift when macros and tone are respected.",
    "Scale safely: expansion playbook, deflection routing, and QA sampling keep quality high as coverage grows."
  ],
  "faq": [
    {
      "question": "How does the copilot respect our existing macros?",
      "answer": "We index your macros, detect the best match per ticket, and draft an extension when context requires. The macro ID and citations are shown in the composer; agents can insert, edit, or escalate."
    },
    {
      "question": "Will this increase risk or cause off-brand replies?",
      "answer": "No. Drafts never auto-send, tone profiles are tuned with QA, and every prompt and draft is logged. QA can sample by queue or agent, and disallowed phrases are enforced."
    },
    {
      "question": "What if our KB is out of date?",
      "answer": "The retrieval layer blends KB with solved tickets and release notes. Low-confidence drafts trigger a KB task and escalate to Tier 2 for complex cases."
    }
  ],
  "business_impact_evidence": {
    "organization_profile": "Global B2B SaaS with 120 agents across NA/EU, Zendesk for tickets, ServiceNow for incidents, 24/5 coverage.",
    "before_state": "Agents copy/pasted macros and KB snippets; AHT at 13m in Billing and 17m in SSO; CSAT at 82.1; FRT lagging on P2s.",
    "after_state": "In-line drafts extended macros with product/version context; AHT dropped to 9.4m (Billing) and 12.2m (SSO); CSAT reached 86.3; FRT improved by 36% on P2s.",
    "metrics": [
      "AHT reduction: 28% across pilot queues",
      "CSAT lift: +4.2 points",
      "First Response Time: 36% faster on P2 incidents",
      "Tier 2 escalations: -19% in SSO"
    ],
    "governance": "Security approved due to RBAC mapped to Zendesk/ServiceNow groups via SSO, prompt logging with ticket IDs, PII redaction, EU data residency enforcement, human-in-the-loop approvals, and a strict policy of never training models on client data."
  },
  "summary": "Support leaders: embed a governed copilot in Zendesk/ServiceNow in 30 days—draft replies, surface fixes, and respect macros. Faster resolutions, higher CSAT."
}