What it means
- New AI products and changes to existing ones pass through defined gates before reaching customers: idea, risk assessment, build or buy, testing, approval, release.
- AI-specific changes need control too: retraining, new training data, prompt or system instruction edits, switching model version or vendor, changing thresholds, adding tools to an agent.
- Each gate has an approver and evidence. High risk changes need sign-off from someone independent of the builders.
- There is a way to roll back or withdraw a model quickly if it misbehaves.
Why it matters
- RBI FREE-AI says every AI-enabled product should go through product approval with AI-specific checks by a team independent of developers.
- Many AI incidents come from "small" changes: a prompt edit or a vendor's silent model upgrade.
- Proposed legislation (the AI Ethics and Accountability Bill 2025) floats withdraw-and-fix powers; a tested rollback path is good preparation.
Checklist
- Define AI product approval gates and who approves at each, scaled by tier.
- Add AI change types to the change management policy (retrain, data change, prompt change, model or vendor version change, threshold change).
- Require risk reassessment (SC-AI-005 AI risk assessment) and retesting (SC-AI-015 Pre-deployment testing and validation) for material changes.
- Version-control prompts, configurations and models; record which version is live.
- Require vendor notice of model changes in contracts and treat them as changes.
- Keep a documented rollback plan and test it.
- Update model and system cards on each change.
Evidence examples
- Product approval record for the AI-based collections tool, with sign-offs from risk, compliance and an independent reviewer.
- Change tickets for model retraining with linked test results and approval.
- Git history or prompt registry showing versioned system prompts.
- Rollback test record (date, time to roll back, result).
Common pitfalls
- Prompt changes made directly in production with no ticket.
- Vendor model upgrades accepted automatically.
- Approval gate run by the build team alone.
- No rollback path because the old model version was deleted.
Satisfies
- VoluntaryISO/IEC 42001
- BindingEU AI Act
- SupervisoryPRA SS1/23
- VoluntaryNIST AI RMF and GenAI Profile
- SupervisoryNYDFS AI Cybersecurity Letter 2024
- BindingKorea AI Basic Act
- SupervisoryOSFI E-23
- VoluntaryRBI FREE-AI
Also relevant to
- AI (Ethics and Accountability) Bill 2025no numbered obligations; mapped in its practical notes (withdraw-and-fix)
Standards crosswalk
Clause references only. Check the standard itself for the exact wording.
ISO/IEC 42001
- 6.3 Planning of changes
- A.6.1.3 Processes for responsible design and development of AI systems
- A.6.2.5 AI system deployment
ISO/IEC 27001
- A.8.31 Separation of development, test and production environments
- A.8.32 Change management
SOC 2
- CC3.4
- CC8.1
Key takeaways
- Treat prompts, data, thresholds and vendor model versions as changes that need approval.
- Strong reuse from ISO 27001 change management (A.8.32, A.8.31) and SOC 2 CC8.1; add AI change types and an independent approval for high risk.
- Always keep a tested rollback path.