AI Center of Excellence: 30‑Day Plan, Champions, Metrics

```yaml
version: "1.3"
artifact: "ai_coe_enablement_playbook"
owners:
  coe_owner: "vp_operations@company.com"
  legal_partner: "associate_gc@company.com"
  security_partner: "dir_sec_arch@company.com"
  data_platform: "head_data_platform@company.com"
regions:
  allowed: ["us-east-1", "eu-west-1"]
  residency_required: true
communication:
  office_hours:
    - domain: "AP Exceptions"
      owner: "ap_champion@company.com"
      schedule: "Wednesdays 10:00–11:00 ET"
      sla_response_hours: 12
    - domain: "Support Triage"
      owner: "cs_champion@company.com"
      schedule: "Tuesdays 15:00–16:00 GMT"
      sla_response_hours: 12
  slack_channel: "#ai-coe-office-hours"
champions:
  quota_per_bu: 2
  capacity_commitment_percent: 12
  selection_criteria:
    - "Respected SME in domain backlog (AP, CS, ITSM)"
    - "Comfortable with SOPs and UAT"
  training_syllabus:
    - "Safe prompting + failure modes"
    - "Human-in-the-loop design"
    - "Telemetry + ROI reporting"
intake:
  form_fields:
    - name: "use_case_name"
    - name: "system_of_record"  # e.g., ServiceNow, Zendesk, Salesforce
    - name: "baseline_metric"   # hours/week, SLA, error_rate
    - name: "data_elements"     # PII? PHI? PCI?
    - name: "expected_outcome"  # e.g., -20% handle time
  risk_tier:
    rules:
      - tier: "low"
        criteria: ["no PII", "read-only", "internal users"]
        approvals: ["coe_owner"]
      - tier: "medium"
        criteria: ["internal PII", "write actions behind human review"]
        approvals: ["coe_owner", "security_partner"]
      - tier: "high"
        criteria: ["external data movement", "financial impact > $25k", "PII cross-border"]
        approvals: ["coe_owner", "security_partner", "legal_partner"]
  slos:
    time_to_triage_hours: 24
    time_to_approval_days:
      low: 2
      medium: 5
      high: 7
pilot:
  evaluation_gates:
    min_eval_precision: 0.88
    min_eval_recall: 0.80
    hallucination_max_rate: 0.02
    human_review_required: ["high"]
  success_metrics:
    wau_target_pct_targeted_roles: 0.55
    hours_returned_per_week: 150
    sla_delta_target_pct: -0.15  # 15% reduction vs baseline
  rollout_steps:
    - step: "sandbox"
      owner: "data_platform"
      exit_criteria: ["eval_set >= 200 cases", "precision/recall thresholds met"]
    - step: "limited_prod"
      owner: "coe_owner"
      exit_criteria: ["WAU >= 40%", "no sev-1 incidents in 2 weeks"]
    - step: "scale"
      owner: "business_unit_owner"
      exit_criteria: ["training complete", "SLA delta sustained 4 weeks"]
telemetry:
  stores:
    - "snowflake.telemetry.ai_usage"
    - "snowflake.telemetry.prompt_logs"
  metrics:
    - "wau_dau_ratio"
    - "hours_returned"
    - "sla_delta"
    - "error_rate"
  observability:
    owner: "head_data_platform@company.com"
    alerts:
      - name: "waudau_drop"
        threshold: 0.35
        action: "page coe_owner"
      - name: "error_rate_spike"
        threshold: 0.05
        action: "rollback_to_human_only"
governance:
  rbac_roles: ["operator", "champion", "reviewer", "approver"]
  prompt_logging: true
  prompt_log_retention_days: 365
  decision_ledger_required: true
  never_train_on_client_data: true
  pii_redaction_enabled: true
  fallback_procedure: "documented_rollback_sop_v2"
deployment:
  runtime: "aws_vpc_private_link"
  model_endpoints: ["bedrock.claude", "azure.openai.gpt4o"]
  vector_db: "aurora_pgvector"
reporting:
  weekly_ops_brief:
    recipients: ["ops_leadership@company.com", "legal_security@company.com"]
    contents: ["adoption_trends", "hours_returned", "sla_deltas", "incidents"]
  monthly_governance_review:
    chair: "associate_gc@company.com"
    evidence: ["prompt_logs", "decision_ledger", "rbac_audit_export"]
```

{
  "title": "AI Center of Excellence: 30‑Day Plan, Champions, Metrics",
  "published_date": "2025-11-16",
  "author": {
    "name": "David Kim",
    "role": "Enablement Director",
    "entity": "DeepSpeed AI"
  },
  "core_concept": "AI Adoption and Enablement",
  "key_takeaways": [
    "Stand up an AI CoE in 30 days using a champions network, scheduled office hours, and a measurable intake-to-production pipeline.",
    "Instrument adoption and ROI from day one: WAU/DAU, hours returned, SLA deltas, and error rates in Snowflake or BigQuery.",
    "Governance is non-negotiable: RBAC via IdP, prompt logs, data residency, and human-in-the-loop for high-risk flows.",
    "Publish an enablement playbook YAML that defines roles, approval steps, SLOs, and evaluation thresholds—so legal says yes and teams ship faster."
  ],
  "faq": [
    {
      "question": "Where should the CoE sit—IT or Operations?",
      "answer": "Operations should own delivery and backlog. IT/Data Platform support integrations and observability. Legal/Security co‑design controls. The CoE succeeds when it is measured on operational KPIs (SLA, hours returned), not just models shipped."
    },
    {
      "question": "How do we avoid office hours becoming a support queue?",
      "answer": "Publish SLAs and scope. Office hours focus on enablement (intake quality, evaluation gates, safety patterns). Triage production incidents through existing ITSM flows with clear ownership."
    },
    {
      "question": "What’s the fastest first pilot?",
      "answer": "Pick a workflow with grounded data and clear owners: AP invoice exception triage in ServiceNow, or support ticket summarization/routing in Zendesk. We’ve shipped both in <30 days with measurable time savings and audit-ready logs."
    }
  ],
  "business_impact_evidence": {
    "organization_profile": "Global logistics provider (7,200 employees) using ServiceNow, Zendesk, Salesforce, Snowflake, AWS; multi-region operations in US/EU.",
    "before_state": "Siloed AI experiments with unclear approvals, no shared metrics, and pilots that took ~90 days to production; legal halted two efforts over residency and logging gaps.",
    "after_state": "Ops-led AI CoE with 22 champions across 4 regions; weekly office hours; governed telemetry in Snowflake; 9 pilots in production within 60 days.",
    "metrics": [
      "1,800 hours per quarter returned to frontline teams (validated time studies).",
      "Pilot-to-production lead time cut from 90 to 24 days (73% faster).",
      "SLA breaches down 22% in AP exceptions and L2 support queues.",
      "58% WAU among targeted roles by week 5; 76% by week 8 in two BUs."
    ],
    "governance": "Approved because deployments ran in customer VPC with region locks (US/EU), RBAC via Okta, prompt logs retained 365 days in Snowflake, decision ledger for approvals, human‑in‑the‑loop on high‑risk actions, and we never train on client data."
  },
  "summary": "Stand up an AI CoE in 30 days: champions, office hours, adoption metrics—governed with RBAC, logs, and data residency. Faster pilots, fewer SLA misses."
}