AI RELEASE READINESS ONE-PAGER (COPY/PASTE)

Feature name:
Owner (accountable for outcomes in production):
Builder team (ships code/workflow):
Gate (security/privacy/legal + AI reliability):
On-call contact / escalation channel:

1) USER IMPACT + SCOPE
- What user job does this do?
- Read-only, draft-only, or can it take actions?
- What is the worst credible failure (one sentence, no euphemisms)?

2) DATA BOUNDARIES
- Data sources used (systems + tables/indices):
- Data sinks (where outputs are stored or sent):
- Sensitive data included? (PII, financial, health, credentials, internal secrets)
- Retention policy for prompts/outputs/traces:
- Access controls: who can view logs and why?

3) ACTION BOUNDARIES (TOOLS)
- Tool allowlist (exact tool names):
- Disallowed actions (explicit):
- Human approvals required for: money movement / outbound comms / data deletion / code merges (circle what applies)
- Sandbox vs production controls:

4) EVALUATION GATES
- What does “correct” mean for this use case?
- Eval dataset location (link):
- Regression cadence (every deploy? daily? weekly?):
- Failure policy: what happens if evals regress?

5) MONITORING + AUDIT
- What is logged (prompt, retrieved docs, tool calls, output, user action)?
- Where are traces stored?
- Alerts: what triggers a page vs a ticket?
- Manual review queue (if any): who owns it and SLA?

6) ROLLBACK / KILL SWITCH
- Kill switch location (config/flag name):
- Default safe mode (read-only, draft-only, disable tools, route to human):
- Who can flip the switch (names/roles):
- Customer comms template link:

7) SHIP DECISION
Gate approval (name/date):
Owner sign-off (name/date):
Launch plan (gradual rollout? cohort? internal first?):

Rule: If you cannot name who can stop the system within minutes, you are not ready to ship autonomy.