What it means
- Someone who did not build the model checks it: its design, data, assumptions, testing and limits.
- Can be an internal model risk or validation team, or an external firm for high risk or complex models.
- Covers bought models too: the validator reviews vendor evidence and runs independent checks on your use case.
- Happens before go-live for medium and high risk systems, and again periodically or after major changes.
Why it matters
- RBI FREE-AI expects models validated before and during use and recommends independent third-party audits for high risk or complex AI.
- RBI's draft model risk management guidance (June 2026) points the same way for AI and ML models.
- Builders mark their own homework too kindly; independence catches blind spots.
Checklist
- Define who validates (internal team reporting outside the build line, or an external firm) and when external validation is required (for example all high risk customer decisioning models).
- Write a validation standard: scope, conceptual soundness, data review, performance and fairness replication, limitation review, findings rating.
- Validate before go-live for medium and high risk; revalidate at a set frequency (for example annually for high risk) and after material change.
- Track findings to closure with owners and due dates.
- Block go-live on unresolved critical findings unless the AI committee formally accepts the risk.
- For external validators, check their independence and competence.
Evidence examples
- Independent validation report for the credit scoring model with findings and ratings.
- Validation findings tracker showing closure dates.
- Engagement letter with an external validator stating scope and independence.
- Org chart showing the validation team reports outside the data science line.
- AI committee minutes accepting a residual finding with conditions.
Common pitfalls
- "Independent" validator sitting in the same team and reporting to the same manager.
- Validation limited to rerunning the developer's notebook.
- Findings logged but never closed.
- Bought models never validated because "the vendor already tested it".
Satisfies
- SupervisoryPRA SS1/23
- VoluntaryNIST AI RMF and GenAI Profile
- SupervisoryOSFI E-23
- VoluntaryRBI FREE-AI
Also relevant to
- ICMR AI in Healthcare Guidelinesno numbered obligations; mapped in its practical notes (clinical validation)
Standards crosswalk
Clause references only. Check the standard itself for the exact wording.
ISO/IEC 42001
- A.6.2.4 AI system verification and validation
ISO/IEC 27001
- A.5.3 Segregation of duties
- A.5.35 Independent review of information security
SOC 2
- CC4.1
Key takeaways
- The builder should never be the only checker for medium and high risk models.
- Partial reuse from ISO 27001: segregation of duties (A.5.3) and independent review (A.5.35) give the pattern, but model validation content is new.
- Track validation findings to closure and block release on critical ones.