Gulf banking AI rules, side by side

A mandatory floor of three categories at §9.7. Two track Annex III closely enough that an EU mapping largely carries over.

No risk tiering binding on banks was identified as of August 31, 2026. The draft Responsible AI Policy is reported to propose four levels (critical, high, limited and low), a structure close enough to the EU model that an existing EU classification would largely survive the mapping.

No statutory list. The institution rates the risk of each AI system itself under §8(d), and the inventory duty under §2(f) covers all AI models and systems regardless of rating.

The full Register goes to the QCB annually and on request, with the high-risk criteria used and a risk and impact assessment.

No registration duty binding on banks was identified as of August 31, 2026. The draft policy is reported to propose mandatory registration for certain AI applications, which would be a new duty for private institutions rather than a transfer of EU work.

Neither regime requires it. The inventory duty under §2(f) is internal, and the Guidance Note does not require it to be filed with the CBUAE on any cycle. This dimension has been assessed and found empty on both sides.

In force since 4 September 2024 under §1, roughly three years ahead of the EU timetable.

Neither identified as of August 31, 2026. A search of the SAMA Rulebook and of English-language secondary reporting found no AI-specific instrument binding on banks and no AI-specific compliance date. SDAIA has issued AI-specific guidance, but it was not confirmed against primary Arabic-language sources for this entry and is not addressed to banks. What does bind are the data protection and cybersecurity obligations reaching banks through SAMA Circular 43045328, in force since 2021 and examinable now.

Guidance is live now and shapes what examiners ask in the current supervision cycle. There is no transition period because there is no new binding rule to transition to.

§9.7.3 makes processing sensitive personal information a mandatory high-risk trigger in its own right, so systems an EU mapping leaves outside the perimeter fall inside it here.

No sector-specific AI instrument binding on banks was identified as of August 31, 2026, in a search of the SAMA Rulebook and of English-language secondary reporting. Requirements arrive through data protection and cybersecurity frameworks that were not written with AI systems in mind.

Four separate QCB approval gates: training, validation and testing results for any new high-risk system (§14.1); high-risk AI procurement before contract signature (§11.2); launch as Provider or material modification (§11.1); and any fully autonomous system whatever its risk rating (§13.5.1, §13.6.1).

Neither regime requires it. The Guidance Note contains no approval gate: nothing requires the CBUAE to see or approve an AI system before it is deployed, or before a vendor contract is signed. This dimension has been assessed and found empty on both sides.

§15.1 expects the Provider to supply development, testing and performance data by contract, and §15.2 requires the Entity to make that full range of data available to the QCB.

Annual cybersecurity reviews of procured AI by independent qualified third parties, documented procurement justification, contractual audit and information rights, and an explicit expectation to avoid single-vendor concentration.

§15.7 prescribes the method: test the model on different demographic groups to see whether any group is systematically advantaged or disadvantaged.

Explicit cadence: once a year, and again on any upgrade, material change, or new model.

An inventory of every AI model, system or technology developed or deployed, with name, purpose and risk rating as minimum metadata.

§14.3 is a close analogue: a User that materially changes an AI Model with a view to placing it on the market under its own name or trademark is considered a Provider. §2 adds that a substantially modified system is a new system, ending the previous life cycle.

The register must rate each high-risk vendor system's substitutability as easy, difficult or impossible and name an alternate provider, with contracts carrying data-wipe exit clauses and a QCB audit right over the provider.

Three named operating modes, applied across consumer-facing AI, with the fully autonomous mode confined to low-risk non-material processes. Substantively close to Article 14, applied to a wider set of systems.

Fully autonomous high-risk systems need built-in guard rails the AI cannot override, reviewed on schedule or on external volatility spikes, with limits linked to warning levels or auto-close routines. Drafted with algorithmic trading in view.

Named and testable: attack surface examination, protection against integrity attacks, query attacks and prompt injection, a data loss prevention tool, and content anomaly detection.

Plain-language, accurate disclosures in Arabic and English, with telephone support in the major languages of the UAE, and a duty to check understandability.

§21.2 requires a two-choice process: supply data to alter the system and resubmit, or request human review. The resubmission limb has no EU analogue.

The Personal Data Protection Law, reaching supervised financial institutions through SAMA Circular 43045328. Different text, similar function: it is the binding floor under any AI system processing personal data.

§23 provides a formal route: request a waiver from a specific requirement with a documented business case, subject to QCB approval, recorded with an expiration date.