What it means
- Plan for AI failing, both loudly (outage, vendor API down) and silently (model quietly getting worse, drift, bad update).
- Each critical AI process has a fallback: a backup model, a rules-based process, or a manual process with trained staff.
- You can stop an AI system instantly if it risks significant harm, and keep the business running.
- ISO 27001 reuse: Partial. ISO 27001 continuity, ICT readiness, backup and redundancy controls give the framework. New for AI: silent degradation triggers, model rollback, sampling of AI decisions and LLM vendor concentration.
Why it matters
- RBI FREE-AI asks regulated entities to extend BCP to AI failures, including silent degradation, with fallbacks, drills and sample human checks of AI decisions.
- RBI FREE-AI also expects the ability to shut down an AI system instantly if it risks significant harm.
- Heavy dependence on one LLM provider is a concentration risk; outages and sudden model changes do happen.
Checklist
- Mark AI systems supporting critical services in the inventory and include them in the business impact analysis.
- For each, document the fallback (secondary model or provider, rules engine, manual queue) and who switches over.
- Set recovery time and capacity targets for the fallback (for example manual processing for 48 hours).
- Build the kill switch into the BCP: who decides, how fast, what customers see (SC-AI-026 Human oversight and override).
- Define silent-failure triggers from monitoring (drift, error rate, complaint spikes) that invoke BCP (SC-AI-020 Performance monitoring and drift detection).
- Sample human checks of AI decisions on a regular schedule to catch silent degradation.
- Back up model artefacts, prompts, configurations and vector indexes so a known-good version can be restored.
- Assess LLM vendor concentration and exit options (SC-AI-007 Third-party and vendor AI due diligence).
- Drill the fallback at least yearly; record results and fixes.
Evidence examples
- BCP and BIA updated with AI systems and fallbacks.
- Fallback runbooks and switch-over decision authority.
- Drill report showing time to switch and issues found.
- Sample human check records of AI decisions.
- Backup records of model artefacts and configuration.
- Vendor concentration assessment for LLM providers.
Common pitfalls
- Manual fallback exists on paper but the team that knew the manual process has moved on.
- BCP only covers outages, not a model that keeps running but gives wrong answers.
- Vendor silently updates the model; no pinned version and no rollback.
- Kill switch with no plan for what customers experience after it is pulled.
Satisfies
- SupervisoryPRA SS1/23
- VoluntaryNIST AI RMF and GenAI Profile
- SupervisoryNYDFS AI Cybersecurity Letter 2024
- SupervisoryOSFI E-23
- VoluntaryRBI FREE-AI
Standards crosswalk
Clause references only. Check the standard itself for the exact wording.
ISO/IEC 42001
- 6.1.3 AI risk treatment
- A.6.2.6 operation and monitoring
ISO/IEC 27001
- A.5.29 information security during disruption
- A.5.30 ICT readiness for business continuity
- A.8.13 information backup
- A.8.14 redundancy of information processing facilities
SOC 2
- A1.1
- A1.2
- A1.3
- CC7.5
- CC9.1
Key takeaways
- Plan for silent failure, not just outages.
- Every critical AI process needs a tested fallback and a tested kill switch.
- Reuse your ISO 27001 BCP; add AI-specific triggers and drills.