AI audit readiness: a practical checklist for CISOs

If you're a CISO in 2026, you don't start with "Let's book an AI audit." You start with "Are we even ready to be audited?"

This is your AI audit readiness checklist. Not theory. Not hype. Just the minimum foundations you need before inviting an auditor, regulator, or your board to look under the hood.

1. Define why you want an AI audit (and what type)

If you don't define the scope and purpose upfront, the AI audit will define it for you. You probably won't like the result.

Start by clarifying which type of AI audit you're aiming for: an AI management system audit (AIMS audit) aligned to ISO/IEC 42001:2023 and NIST AI RMF; a model risk management / ML model validation audit focused on individual models and SR 11-7–style governance; a generative AI security / LLM red-team audit mapped to the OWASP GenAI LLM Top 10 2026 and OWASP Agentic Top 10; or a regulatory-focused AI audit covering the EU AI Act, GDPR, and sector laws.

Your checklist before booking an auditor:

• Write a one-page AI audit charter: purpose, scope, success criteria, and main frameworks (e.g. ISO 42001 and EU AI Act mapping, ISO 42001 vs NIST AI RMF).
• Decide if this is a responsible AI audit, ethical AI audit, or a narrower security or compliance review.
• Decide whether you're targeting external certification (e.g. ISO 42001, ISO 27001, SOC 2 AI/ML audits, CSA STAR for AI), or an internal baseline.

Skip this step, and the audit becomes a fishing expedition. Fishing expeditions rarely end well in front of a board.

2. Build a real AI system inventory (no more "shadow AI")

You cannot audit what you cannot see. CISOs already know this from shadow IT. Now repeat it for shadow AI.

You want to answer, on one page:

"These are our AI systems.
These are the ones in scope for this AI audit.
This is who owns the risk."

Without that, the rest of the checklist doesn't matter.

Readiness checklist:

• Create a central AI system inventory (spreadsheet, GRC tool, or dedicated register) and assign an owner to maintain it.
• For each AI system, record: business owner, AI risk owner, purpose, criticality, data sensitivity, and interaction level.
• Document technical details: model types, training and inference data sources, deployment locations, and key integrations.
• Map AI supply chain dependencies: third-party models, APIs, plugins, tool connectors, and vector stores.
• Run a shadow AI discovery exercise (surveys, access logs, SaaS inventories) to surface unapproved tools and DIY integrations.
• Review vendor contracts and security questionnaires to identify third-party AI risk assessment gaps for partners with AI features.
• Define which systems are in scope for your upcoming AI audit and ensure their owners are aware and engaged.

3. Stand up an AI governance and control framework

An AI audit is not just a technology review. It tests whether you have an AI control framework that people actually follow.

Anchor your approach in ISO/IEC 42001:2023 (38 Annex A controls across A.2–A.10, plus an AI Statement of Applicability and risk treatment plan) and the NIST AI Risk Management Framework (Govern, Map, Measure, Manage), including the NIST AI 600-1 Generative AI Profile for LLM use-cases.

Your readiness checklist:

• Document a single AI governance policy referencing ISO 42001 and NIST AI RMF.
• Define a simple AI governance maturity model (e.g. initial, repeatable, optimised).
• Establish an AI governance officer or committee, an AI audit committee or AI ethics board, and clear AI risk owner roles for major systems.
• Decide how you'll handle AI bias audits, AI explainability / XAI audits, and consistent AI logging and monitoring to create an AI audit trail.

If governance is still a slide deck, you're not ready. The auditor will spot that in the first hour.

4. Align with the EU AI Act (even if you're not in the EU)

You might think the EU AI Act doesn't apply to you. But its structure is quickly becoming a global template. See how it applies in Switzerland in my previous article here.

An AI audit in 2026 will often map your controls to key EU AI Act provisions: Article 5 (prohibited practices), Article 6 and Annex III (high‑risk classification), Article 9 (risk management), Article 12 (audit logging), Article 14 (human oversight), and Article 50 (transparency). Annex IV sets out the required AI technical documentation, while conformity assessment, notified bodies, and CE marking apply to high‑risk systems.

Readiness checklist:

• Classify all systems using Article 6 and Annex III criteria to identify potential high-risk AI.
• For potential high-risk systems, prepare a risk management plan (Article 9), an AI incident log, a post-market monitoring plan, and draft Annex IV-style technical documentation.
• For systems affecting individuals' rights, build a Fundamental Rights Impact Assessment (FRIA) template.

If an auditor asks "Show me your Annex IV-like documentation and FRIA for this system" and you reach for a blank notebook, you're not ready.

5. Integrate data protection: GDPR and beyond

Most AI audits run into data-protection issues by the second meeting. Address them from the outset.

Build key GDPR concepts into the audit scope, especially Article 22, which regulates solely automated decisions that significantly affect individuals and requires appropriate safeguards, including meaningful human intervention and a genuine opportunity to challenge the outcome.

Uber’s experience is a useful warning: the Amsterdam Court of Appeal found that several driver deactivations lacked meaningful human involvement and that Uber failed to provide adequate information about its automated decision-making. The Dutch regulator has since imposed an €825 million fine on Uber over automated driver deactivations.

Also assess whether the AI system requires a data protection impact assessment (DPIA), and align it where relevant with broader algorithmic impact assessments. Finally, document the lawful basis for collecting and using training data, prompts, outputs, monitoring data, and system logs.

Readiness checklist:

• Update your DPIA templates to explicitly handle LLMs, agentic AI, cross-border data flows, and training data retention/reuse.
• Decide when you require AI bias audits and XAI audits for automated decisions.
• Map your AI risk register to your privacy risk register so AI and privacy risks are managed together.

If your privacy team only hears about an AI system at go-live, you have a governance issue, not a tooling issue.

6. Address sector-specific AI obligations

A generic AI audit is not enough in regulated sectors. Different regulators have their own expectations.

Plan for domain-specific rules such as AI in medical devices (EU MDR / FDA AI/ML SaMD), AI in financial services (SR 11-7, EBA AI guidelines, MAS FEAT principles), and AI in hiring (NYC Local Law 144, Illinois AI Video Interview Act).

Readiness checklist:

• For each regulated domain, document applicable regulations, guidance, and any additional testing, documentation, and oversight requirements.
• Integrate these sector obligations into your AI management system and AI control framework, not as an afterthought.

Failing to connect general AI governance with sector rules is how you pass one audit and fail the one that matters.

7. Harden your LLM and agentic applications

Today's AI incidents are less about "bad math" and more about bad integration. This is where OWASP and GenAI security come in.

Use the OWASP Top 10 for LLM Applications 2026 (LLM01–LLM10: prompt injection, sensitive info disclosure, supply chain, data/model poisoning, improper output handling, excessive agency, system prompt leakage, vector/embedding weaknesses, misinformation, unbounded consumption) and the OWASP Top 10 for Agentic Applications 2026 (ASI01–ASI10) as your core security baseline.

Readiness checklist:

• Build an OWASP GenAI LLM Top 10 2026 audit checklist and use it for LLM threat modeling, LLM red-team testing, and generative AI security audits.
• For autonomous agents, apply OWASP Agentic Top 10 security controls, limit tools and permissions, and run an agentic AI security audit for systems with external actions.
• Implement recurring LLM red-team audits, guardrails for excessive agency, and clear boundaries between test and production environments.

If your LLM and agents can "do anything anywhere," an AI audit is the least of your worries.

8. Align AI security with ISO 27001 and SOC 2

AI doesn't live outside your security program. Good auditors will check how AI fits into familiar frameworks.

Be ready to show how you extend ISO/IEC 27001 (Annex A controls for AI workloads, plus an AI risk assessment addendum) and SOC 2 (Trust Services Criteria applied to AI, an AI system description, and mapping to Security, Confidentiality, and Processing Integrity). Maintain a cross-map between ISO 42001, NIST AI RMF, ISO 27001, SOC 2, and OWASP LLM/Agentic Top 10.

Readiness checklist:

• Update your ISMS to explicitly cover AI instead of letting it live on a separate island.
• Maintain a cross-map between ISO 42001 Annex A, NIST AI RMF, ISO 27001 Annex A, SOC 2 criteria, and OWASP LLM/Agentic Top 10.
• For cloud-based or agentic AI, consider CSA STAR for AI certification, the STAR for AI Agentic Certification Scheme, and track relevant STAR Registry ISO 42001 certificates.

The goal is simple: one coherent control story. Not five disconnected frameworks.

9. Clarify AI roles, competencies and accountability

Auditors don't just look at controls. They look at who owns them.

Define key roles such as AI auditor / AIMS auditor / ISO 42006 auditor, AI governance officer, model risk manager, AI risk owner, and an AI audit committee or AI ethics board. Note that ISO/IEC 42006:2025 specifies requirements for bodies providing audit and certification of AI management systems, so expect questions about your certifier's accreditation and auditor competence.

Readiness checklist:

• Document a RACI for AI risk assessments, change approvals, incident response, and audit findings/remediations.
• Define minimum competence requirements (AI/ML fundamentals, security and privacy, regulatory understanding).
• Provide targeted training for teams who will be interviewed in an AI audit and for management who will sign off on findings and risk acceptance.

If "who owns this AI risk?" leads to silence in the room, you're not ready.

10. Prepare audit-ready artefacts and run a mock audit

An AI audit lives and dies by evidence. Good intentions don't count. You need tangible artefacts ready for review.

Readiness checklist:

• Gather core governance artefacts: AI policy, AI risk treatment plan, AI Statement of Applicability, AI risk register, AI system inventory, and AI impact assessment records.
• Compile system-level documentation: model cards, data sheets, Annex IV-style technical documentation, DPIAs, FRIAs, training data lineage, and bias/robustness test results.
• Assemble security and testing evidence: LLM red-team reports, OWASP LLM threat model outputs, AI audit trail and logging configurations, human oversight logs, and incident/exception records.
• Conduct a mock AI audit using internal audit or a friendly external party, with a fixed timeline, defined scope, formal findings, and a remediation plan.

If your artefacts are scattered across ten shared drives and a wiki no one updates, you're not ready.

11. Plan for continuous monitoring and post-audit life

Passing an AI audit once is not the endgame. Regulators and boards increasingly expect continuous AI governance, not point-in-time box-ticking.

Build in ongoing monitoring of your AI risk register and AI system inventory, periodic re-testing against OWASP LLM/Agentic Top 10, and continuous AI logging and monitoring with clear alerting and escalation. Implement versioned change management for models, prompts, tools, and agent configurations, and run at least annual internal AI audits with management review minutes, nonconformity records, and corrective actions.

Readiness checklist:

• Define how often you will refresh your AI governance maturity model, re-run mock AI audits, and update your AI Statement of Applicability and risk treatment decisions.
• Ensure your AI audit committee or equivalent body sees incident trends, reviews major model changes and new high-risk use-cases, and challenges whether your controls are keeping pace with AI adoption.

If your AI governance stops the day after the audit report lands, you've built a compliance exercise, not a resilient AI programme.

Use this checklist as your baseline. Adapt it to your industry, risk appetite, and regulatory footprint. But don't skip the boring parts: inventory, evidence, roles, and continuous monitoring. Those are exactly where AI audits are won or lost.