# AI Release Gate Checklist
Executive Summary
This checklist turns AI security requirements into an engineering release gate. It covers model and provider changes, prompt changes, retrieval sources, agent tools, logging, privacy, approval, rollback, and signoff.
The goal is simple: high-risk AI changes should not ship because everyone assumed someone else checked them.
Public sample notice
Sample release decision
No-go for Northstar Support Cloud / Customer Support Copilot source expansion. Retrieval authorization negative tests and approval context bundles are incomplete.
Release Gate Snapshot
The release gate is where governance becomes real
Change categories
AI change categories
| Change category | Risk | Required checks |
|---|---|---|
| Model or provider change | High | provider approval, data-use statement, routing policy, fallback behavior |
| Prompt change | Medium | prompt diff, prompt injection test, instruction isolation, rollback version |
| Retrieval source change | Critical | ACL metadata, tenant isolation, chunk metadata, reranker safety |
| Tool policy change | Critical | permission matrix, action classes, approval bundle, trace, rollback |
| Logging or trace change | High | schema review, retention, access control, redaction, reconstruction |
| Customer-facing answer change | Medium | answer bank update, evidence link, legal review, freshness |
AI Release Gate Checklist
The checklist tracks required controls, owners, evidence, signoff, and go/no-go decisions for AI-related changes.
Required checks
Required checks before AI release
Release-blocking findings
Release Gate Findings
Retrieval tests block source expansion
The release expands retrieval source coverage, but negative authorization tests are incomplete. This should block release until tests pass.
Approval context blocks sensitive action expansion
Sensitive actions require a meaningful approval context. A confirmation click is not enough for customer-visible or state-changing tool use.
Customer-facing answers need refresh
If a release changes model routes, retrieval behavior, tool authority, logging, or retention, the answer bank and evidence pack must be updated.
Signoff model
AI release signoff model
| Role | Required for | Decision |
|---|---|---|
| Product Security | all AI releases | required |
| AI Platform Engineering | prompts, model routes, tool policy | required |
| Search Platform | retrieval, indexing, reranking | required |
| Privacy Engineering | retention and customer data processing | conditional |
| Legal | provider claims and customer-facing answer changes | conditional |
Go/no-go model
Critical unknowns default to no-go. High unknowns default to conditional-go. Approved releases must include evidence, owners, rollback, and signoff.
Engineering use
How engineering should use the gate
| Moment | Action |
|---|---|
| Planning | identify AI change categories and required checks |
| Pre-merge | complete prompt, retrieval, provider, tool, logging, and privacy evidence |
| Release review | confirm owners and signoff |
| Post-release | monitor traces, exceptions, incidents, and buyer answer drift |
| Rollback | execute rollback if traces, permissions, or output behavior violate gate assumptions |
Related artifact: RAG Security Test Plan
The RAG test plan supplies release-gate evidence for retrieval changes.
Related artifact: Agent Tool Permission Matrix
The permission matrix supplies release-gate evidence for agent and tool-policy changes.
Related artifact: AI Security Operating Model Blueprint
The operating model defines where the release gate lives and who owns it.