CFO Playbook: Natural‑Language Query in Snowflake/Power BI Without Data Leakage — A 30‑Day, Audit‑Ready Rollout

```yaml
version: 1.3
owner:
  business: CFO (Finance)
  data: Director, Data Engineering
  security: Head of Internal Audit
regions:
  primary: us-east-1
  residency: true  # data and prompts remain in-region
stack:
  warehouse: snowflake
  bi_semantic_layer: power_bi
  idp: azure_ad
nlq_agent:
  model_provider: enterprise_llm_vpc  # on‑prem/VPC; no training on client data
  compile_target: semantic_views_only  # block direct table access
  allowed_sql:
    - SELECT
    - WITH
    - WINDOW_FUNCTIONS
  blocked_sql:
    - INSERT
    - UPDATE
    - DELETE
    - MERGE
  join_policies:
    approved_dimensions:
      - dim_account
      - dim_region
      - dim_product
    forbidden_joins:
      - workday_compensation -> sales_opportunity  # requires approval
  row_limits:
    default: 100000
    max: 500000
  confidence_thresholds:
    auto_answer: 0.82
    require_human_review: 0.65
    block_below: 0.50
  ambiguity_resolution:
    route_to: fpa_teams_channel
    sla_minutes: 30
security:
  rbac_source: snowflake_roles
  rls: inherit_from_semantic_layer
  masking_policies:
    - name: mask_pii
      applies_to:
        - ssn
        - bank_account
        - personal_email
      method: partial_hash
  pii_redaction:
    enabled: true
    patterns:
      - '\\b\\d{3}-\\d{2}-\\d{4}\\b'  # SSN
      - 'acct\\d{6,}'
observability:
  prompt_logging: enabled
  sql_logging: enabled
  lineage_capture: enabled
  retention_days: 365
  decision_ledger:
    materiality_threshold:
      metric_impact_pct: 1.0
      confidence_min: 0.82
    storage: s3://finance-analytics-evidence/decision-ledger/
approvals:
  rules:
    - name: workday_join
      when:
        join: workday_* -> *
      approvers:
        - Controller
        - Internal_Audit
      sla_minutes: 15
      channel: teams
    - name: attribute_exposure_payroll
      when:
        fields: [salary, bonus]
      approvers: [Controller]
      sla_minutes: 10
slo:
  answer_latency_p50_ms: 2500
  answer_latency_p95_ms: 8000
  availability: 99.5
telemetry:
  metrics:
    - unauthorized_attempt_count
    - redaction_applied_count
    - approval_request_latency
    - nlq_to_sql_parse_errors
egress_policy:
  external_calls: deny
  training_on_client_data: false
  export:
    csv: deny
    pdf: allow_if_watermarked
```

{
  "title": "CFO Playbook: Natural‑Language Query in Snowflake/Power BI Without Data Leakage — A 30‑Day, Audit‑Ready Rollout",
  "published_date": "2025-11-10",
  "author": {
    "name": "Elena Vasquez",
    "role": "Chief Analytics Officer",
    "entity": "DeepSpeed AI"
  },
  "core_concept": "Executive Intelligence and Analytics",
  "key_takeaways": [
    "Put NLQ agents inside Snowflake/BigQuery/Databricks and your BI semantic layer—no shadow extracts.",
    "Enforce RBAC, row‑level security, redaction, and prompt logging at runtime to satisfy SOX and internal audit.",
    "30‑day motion: Week 1 metric inventory; Weeks 2–3 semantic guardrails; Week 4 board‑grade brief and alerting.",
    "Outcome to repeat: variance root‑cause time cut 72%, with 38% analyst hours returned in month one."
  ],
  "faq": [
    {
      "question": "Will NLQ break our SOX boundary or force re‑certifying reports?",
      "answer": "No. We compile against certified semantic views and inherit RBAC/RLS. The agent cannot access raw tables or modify data. Every answer carries lineage and a decision‑ledger record for audit."
    },
    {
      "question": "How do you prevent leakage to external model providers?",
      "answer": "We deploy on‑prem/VPC model endpoints and disable external egress. Prompts and queries stay in‑region with logging and retention policies you control."
    },
    {
      "question": "What happens when the question is ambiguous or the confidence is low?",
      "answer": "Below the auto‑answer threshold, the request routes to FP&A for clarification with the compiled SQL attached. You set thresholds and SLAs for human‑in‑the‑loop review."
    },
    {
      "question": "How much data modeling is required?",
      "answer": "We reuse your existing BI semantic layer. Week 1 inventories the top questions and maps them to certified datasets and approved joins; only gaps are modeled, not a full rebuild."
    }
  ],
  "business_impact_evidence": {
    "organization_profile": "$800M ARR global SaaS company with Snowflake, Power BI, Salesforce, and Workday; finance team of 45, distributed across US/EU.",
    "before_state": "Executives asked ad‑hoc questions in Slack; FP&A spent days compiling one‑off pulls and reconciling conflicting metrics across decks. Legal blocked prior NLQ pilots due to data egress and lack of logging.",
    "after_state": "NLQ embedded in Power BI against Snowflake semantic views with RBAC/RLS and prompt logging. CFO and VPs ask questions in Teams; answers return with lineage, confidence, and decision ledger entries.",
    "metrics": [
      "Variance root‑cause cycle reduced from 3.5 days to 12 hours (−72%).",
      "38% analyst hours returned in month one by eliminating ad‑hoc data pulls.",
      "0 audit findings; 100% of prompts and SQL compiled/logged; all queries executed in‑region."
    ],
    "governance": "Legal/Security approved because data never left Snowflake/region, RBAC/RLS were inherited, prompts/SQL/answers were logged with retention, sensitive joins required approvals, and the model never trained on client data."
  },
  "summary": "CFOs: ship secure NLQ inside Snowflake/Power BI in 30 days. Faster answers, governed access, zero data egress—RBAC, prompt logs, and residency built in."
}