MODEL GOVERNANCE ONE-PAGER (TEMPLATE)

Purpose
- Workflow name:
- Customer-facing surface(s):
- Business owner (name/role):
- Technical owner (name/role):
- What happens if this workflow is wrong? (1–2 sentences, concrete harm)

Model + Provider
- Provider (OpenAI / Azure OpenAI / Anthropic / Google / AWS Bedrock / Self-hosted):
- Model ID + version/date:
- Why this model (capability, enterprise fit, latency, control):
- Change policy: who can change the model or prompt? How is approval done?

Data Handling
- Inputs (what data types enter the model):
- Sensitive data present? (PII, PHI, credentials, source code, contracts):
- What is sent to the provider (prompts, files, tool outputs, embeddings):
- Retention terms and contract notes (DPA, region constraints):
- Redaction rules before sending:

Permissions + Tooling
- Tools/functions the model can call:
- Permission scope for each tool (least privilege):
- Network/data access constraints (allowlist, sandboxing):
- Prompt injection mitigations (where applicable):

Evaluation (Release Gate)
- Eval suite name:
- What does “pass” mean? (explicit criteria):
- Regression policy (what triggers a block):
- Safety tests included (misuse cases, jailbreak attempts):
- Human review requirements (if any):

Observability + Audit
- Logging: what is logged (inputs/outputs/tool calls) and what is excluded:
- Redaction applied to logs:
- Log retention window:
- Access controls to logs (roles):
- Trace ID approach (how to reconstruct a single workflow execution):

Reliability + Fail-Soft Behavior
- Timeouts and retry policy:
- Rate limits/backpressure behavior:
- What the product does when the model is unavailable (fallback UI / cached answer / human handoff):

Cost Controls
- Budget owner (role):
- Quotas (per tenant / per user / per workflow):
- Caching strategy (if applicable):
- Alerting thresholds (what triggers a page/ticket):

Rollback Plan
- What can be rolled back quickly (model version, prompt, tool access, retrieval corpus):
- How rollback is executed (flag, config, deploy):
- Who is on-call / escalation path:

Approvals
- Product approval (name/date):
- Engineering approval (name/date):
- Security/privacy approval (name/date):
- Legal/compliance approval (name/date, if needed):

Notes
- Known limitations:
- Open questions and owners: