Guide for Auditors of AI Systems in the GCC
Verification required. Regional claims on this page were compiled from secondary sources on 7 October 2026. No primary legal text was read, and each claim is tagged [Reported], [Conflict] or [Unverified] as explained on the section home page. Check the primary text before relying on any of them. This is not legal advice.
For internal audit, external assurance providers and anyone preparing for a regulator review. Regional claims carry the tags used in the GCC AI Regulatory Hub.
1. Set the audit basis honestly
Define the criteria before testing. A GCC AI audit usually combines:
| Criteria source | Examples | Caution |
|---|---|---|
| Binding law | UAE PDPL, DIFC Reg 10, Saudi PDPL, Qatar PDPPL | Read the primary text. This wiki does not reproduce it. |
| Sector regulator | CBUAE note, QCB guideline | Binding status differs; confirm |
| National baseline | NCA, Dubai AI Security Policy, NCSA | Some are voluntary or in consultation |
| Internal policy | Client AI policy | Always in scope |
| Framework | Wiki controls, ISO/IEC 42001, NIST AI RMF | Not a legal standard |
State in the engagement letter which criteria are in scope and which are reference only. Do not report "non-compliant with law X" on evidence of a secondary summary.
2. Scoping steps
- Obtain the AI inventory (T09). If none exists, that is finding one.
- Tier systems (T01). Sample all high-impact systems, and a risk-weighted sample of others.
- Identify jurisdictions per system: users, data subjects, hosting, vendor.
- Identify regulators and any notification or approval duties (T11).
- Agree evidence period and sampling method.
3. Control families and test steps
| Family | Wiki controls | Key test | Evidence |
|---|---|---|---|
| Governance | AI-CTRL-011, AI-CTRL-012, AI-CTRL-037 | Roles named; board reporting; policy approved and current | Policy, minutes, RACI |
| Inventory | AI-CTRL-011 | Inventory completeness test: trace 5 known systems into register and 5 register items into reality | Register, discovery output |
| Privacy and regulatory | AI-CTRL-013, AI-CTRL-017 | Lawful basis recorded; impact assessment done before launch | T02 records |
| Residency | AI-CTRL-007 | Attestation matches actual configuration | Cloud config export, T04 |
| Vendor | AI-CTRL-014 | Due diligence done before contract; sub-processors listed | T03, contract |
| Oversight | AI-CTRL-023, AI-CTRL-025 | Human approval exists and is evidenced; kill switch test within period | Review logs, test record |
| Content safety and reliability | AI-CTRL-039 | Test results current; Arabic included | Test reports |
| Fairness | AI-CTRL-041 | Bias test current; thresholds approved | T10 |
| Incident | AI-CTRL-035 | Playbook; exercise; notice workflow includes regulators | T05, exercise log |
| Audit evidence | AI-CTRL-038 | Evidence retained and retrievable | Evidence register |
Use T08 for a line-by-line checklist and the Evidence Register to track requests.
4. Sampling guidance
| Population | Suggested sample | Note |
|---|---|---|
| High-impact AI systems | All | Small population |
| Medium | Risk-based, minimum 25 percent or 5, whichever is larger | A suggested default, not a standard |
| Decisions for explanation test | 10 per system | Trace to input, output, explanation, review |
| Incidents | All in period | Check notice timing |
These thresholds are practitioner defaults, not a regulatory requirement. Document your own rationale.
5. Techniques
- Re-performance: re-run a bias or accuracy test on a fixed dataset and compare to reported results.
- Configuration inspection: inspect hosting region, logging, retention, model version.
- Walkthrough: follow one decision end to end including human review.
- Interview: owner, engineer, privacy lead, vendor manager.
- Document inspection: approvals dated before go-live.
- Kill-switch observation: witness the test.
6. Common findings
- Inventory incomplete (shadow AI, embedded vendor AI).
- Impact assessment done after launch.
- Residency attestation not matched to configuration (backups or logs elsewhere).
- Human review exists on paper but reviewers approve everything in seconds.
- Bias test performed once, no thresholds.
- English-only testing for Arabic-speaking users.
- Vendor sub-processors unknown.
- No named regulator contact or notification path.
7. Rating and wording
Suggested scale: Effective / Partially effective / Ineffective / Not tested. Word findings as: criteria, condition, cause, effect, recommendation.
Template sentence for secondary-source criteria: "The assessed system was evaluated against [criteria] as summarised in the client's regulatory register dated [date]. The register has not been verified against primary legal text."
8. Report checklist
- Scope, period, criteria and their verification status
- Systems and sample
- Limitations (secondary sources, no legal opinion)
- Findings with evidence references
- Management responses with dates
- Follow-up plan
9. Independence and handling
Keep evidence containing personal data under the same rules as production. Do not copy personal data into external AI tools during the audit.