AI Support Copilot: Single Queue Pilot, Multilingual Scale

# voc_pipeline.yaml
version: 1.3
owners:
  product_owner: "maria.liu@support.example.com"
  qa_lead: "devon.khan@support.example.com"
  data_steward: "jordan.nguyen@it.example.com"
  regional_mgrs:
    - region: "NA"
      owner: "amelia.ross@support.example.com"
    - region: "EU"
      owner: "luc.moreau@support.example.com"
    - region: "LATAM"
      owner: "sofia.garcia@support.example.com"
queue:
  name: "Billing L2"
  platform: "zendesk"
  ticket_forms: [12345, 67890]
  macros_required: true
slo:
  aht_target_seconds: 600
  first_response_minutes: 15
  qa_sample_size_per_locale: 50
  qa_pass_threshold: 0.90
  confidence_gate: 0.78
  csat_delta_min_for_scale: 2.0  # points increase required
  reopen_rate_max_for_scale: 0.07
telemetry:
  prompt_logging: true
  rbac_roles:
    - role: "agent" 
      permissions: ["draft_view", "approve_reply"]
    - role: "qa_reviewer"
      permissions: ["approve_reply", "flag_issue", "publish_feedback"]
    - role: "manager"
      permissions: ["approve_reply", "pause_locale", "promote_locale"]
  data_residency:
    na: "us-east-1"
    eu: "eu-central-1"
    latam: "sa-east-1"
locales:
  - code: "en-NA"
    status: "pilot"
    glossary: ["refund window", "chargeback", "proration"]
    model_allowlist: ["vpc-gpt-4o-na"]
    auto_send: false
  - code: "es-ES"
    status: "candidate"
    glossary: ["devolución", "reclamación bancaria", "prorrateo"]
    model_allowlist: ["vpc-gpt-4o-eu"]
    auto_send: false
    fallback_locale: "en-NA"
  - code: "fr-CA"
    status: "candidate"
    glossary: ["remboursement", "rétrofacturation", "prorata"]
    model_allowlist: ["vpc-gpt-4o-na"]
    auto_send: false
    fallback_locale: "en-NA"
feedback_routes:
  agent_feedback:
    destination: "slack://#copilot-quality"
    fields: ["useful", "accurate", "tone", "notes"]
  csat_comments:
    destination: "s3://support-voc/csat_comments/"
  qa_reviews:
    destination: "servicenow://qa_table"
quality_gates:
  promote_locale_criteria:
    - metric: "confidence"
      op: ">="
      value: 0.80
    - metric: "qa_pass_rate"
      op: ">="
      value: 0.92
    - metric: "csat_delta"
      op: ">="
      value: 2.0
    - metric: "reopen_rate"
      op: "<="
      value: 0.07
  pause_locale_criteria:
    - metric: "confidence"
      op: "<"
      value: 0.70
    - metric: "qa_pass_rate"
      op: "<"
      value: 0.85
approvals:
  change_control:
    required: true
    approvers: ["qa_lead", "regional_mgrs.owner"]
    sla_hours: 24
  incident_response:
    severity_threshold: "medium"
    notify: ["security@company.com", "legal@company.com"]
    evidence_log: "s3://audit/copilot/prompt_logs/"

{
  "title": "AI Support Copilot: Single Queue Pilot, Multilingual Scale",
  "published_date": "2025-11-29",
  "author": {
    "name": "Alex Rivera",
    "role": "Director of AI Experiences",
    "entity": "DeepSpeed AI"
  },
  "core_concept": "AI Copilots and Workflow Assistants",
  "key_takeaways": [
    "Prove impact in one queue first; expand by language and region only after VoC thresholds are met.",
    "Keep agents in control with draft, suggest, and approve states—no auto-send during the pilot.",
    "Instrument governance from day one: prompt logs, RBAC, data residency, and red-team scenarios.",
    "The 30-day motion: Week 1 knowledge audit and voice tuning; Weeks 2–3 retrieval and prototype; Week 4 analytics and expansion plan.",
    "Anchor the business case to one clear metric—e.g., AHT down 18% in the pilot queue—then scale."
  ],
  "faq": [
    {
      "question": "Why start with one queue instead of deploying broadly?",
      "answer": "You’ll prove impact, harden governance, and learn which macros/KB gaps limit quality. One queue gives clean before/after metrics and a safe place to tune tone and retrieval."
    },
    {
      "question": "How do you handle languages with different refund policies?",
      "answer": "We shard retrieval by locale, maintain regional glossaries, and encode policy differences. Confidence gates and QA sampling prevent scale until quality is proven."
    },
    {
      "question": "Can we run this without sending data to a public model?",
      "answer": "Yes. We support VPC or on‑prem endpoints with region pins. Sensitive fields are redacted pre‑prompt, and nothing is used to train models."
    },
    {
      "question": "What does ‘agent-in-the-loop’ look like in the UI?",
      "answer": "Agents see draft replies with sources and a confidence score. They can approve, edit, or flag. Auto‑send is off during pilot, and approvals are logged for audit."
    }
  ],
  "business_impact_evidence": {
    "organization_profile": "Global SaaS platform, 1,100 agents, Zendesk, four languages (EN, ES, FR, DE).",
    "before_state": "Billing L2 AHT at 11m42s, CSAT 82.6, reopen rate 9.1%, Monday backlog spikes +28%.",
    "after_state": "AHT reduced to 9m36s (−18%), CSAT +4.2 points, reopen rate 6.3%, backlog −24% within 30 days of pilot.",
    "metrics": [
      "−18% AHT in pilot queue",
      "+4.2 CSAT points in 30 days",
      "−2.8pp reopen rate",
      "−24% pilot queue backlog"
    ],
    "governance": "Legal/Security approved because all prompts and outputs are logged with ticket/user IDs, RBAC enforced approvals, region-specific data residency (VPC endpoints), and models never trained on client data."
  },
  "summary": "Pilot a support copilot in one queue, capture VoC, and scale across regions and languages in 30 days—governed, auditable, and agent‑in‑the‑loop."
}