What it means
- Internal audit (or a qualified independent function) checks, on a risk-based plan, that AI controls exist and work.
- Scope covers the three layers of each AI system: input data, the model and algorithm, and outputs and decisions.
- High-risk or complex AI gets an independent third-party audit as well.
- The audit framework itself is reviewed regularly.
- ISO 27001 reuse: Partial. Your ISO 27001 internal audit and independent review programme, auditors and finding tracker can be reused. The AI scope (data, models, outputs, fairness, oversight) and AI-competent auditors are new.
Why it matters
- RBI FREE-AI asks for a risk-based AI audit covering data, model and outputs, third-party audits for high-risk or complex AI, checks of stop, pause and unwind mechanisms, an AI BCP and human override, and a review of the audit framework at least every two years.
- NITI Aayog's accountability principle and the India AI Governance Guidelines both point to audits as core.
- ISO/IEC 42001 certification requires a working internal audit programme.
Checklist
- Add AI to the annual internal audit plan, prioritised by AI risk tier (SC-AI-004 AI risk classification and use-case tiering).
- Build an AI audit programme covering: governance, data, model validation, fairness, security, oversight, monitoring, incidents, vendors.
- Include specific tests for kill switch, pause and rollback, AI BCP and human override (SC-AI-026 Human oversight and override, SC-AI-033 Business continuity and fallback for AI).
- Make sure auditors are competent in AI (train them or co-source specialists) and independent of the teams audited.
- Commission an external audit for high-risk or complex AI (SC-AI-016 Independent model validation).
- Track findings with owners and due dates; verify closure.
- Report audit results to the board or audit committee (SC-AI-040 Board reporting and AI risk metrics).
- Review the AI audit framework at least every two years.
Evidence examples
- Annual audit plan showing AI audits and the risk basis for selection.
- AI audit programme and test scripts.
- Audit reports with findings, ratings and management responses.
- Finding tracker with closure evidence.
- Auditor competence records (training, certifications, co-source contracts).
- Board or audit committee minutes noting AI audit results.
- Record of the latest review of the AI audit framework.
Common pitfalls
- Auditing only the AI policy document, not how models actually behave.
- Auditors lacking the skills to challenge data scientists.
- Same team builds and audits the model.
- Findings closed on promise rather than evidence.
Satisfies
- VoluntaryISO/IEC 42001
- SupervisoryPRA SS1/23
- VoluntaryRBI FREE-AI
Also relevant to
- NITI Aayog Responsible AI for Allguidance mapping (audit process; no obligation IDs)
- India AI Governance Guidelinesguidance mapping (algorithmic audits; no obligation IDs)
Standards crosswalk
Clause references only. Check the standard itself for the exact wording.
ISO/IEC 42001
- 9.1 monitoring, measurement, analysis and evaluation
- 9.2 internal audit
- 10.2 nonconformity and corrective action
ISO/IEC 27001
- A.5.35 independent review of information security
- A.5.36 compliance with policies, rules and standards
SOC 2
- CC4.1
- CC4.2
Key takeaways
- Audit data, model and outputs on a risk-based plan.
- Test the safety nets (kill switch, BCP, human override), not just paperwork.
- Reuse the ISO 27001 audit machinery; add AI scope and skills.