Guide

Operational Readiness for AI in UAE Banks and Insurers

The CBUAE's February 2026 Guidance Note places consumer protection at the centre of AI and machine-learning adoption by licensed financial institutions. A bank or insurer must therefore connect each AI use with the people accountable for its consumer outcome, the evidence required before release and the controls that continue after release. A separate AI policy cannot provide that operating discipline on its own.

Start with the consumer decision

The Central Bank of the UAE issued the Guidance Note on the Consumer Protection and Responsible Adoption and Use of Artificial Intelligence and Machine Learning by Licensed Financial Institutions on 11 February 2026. It applies to licensed financial institutions, including insurance providers, and concentrates on uses that can affect consumers. The Guidance supplements rather than replaces applicable laws, regulations and CBUAE directives.

The institution must therefore begin with the consumer process and the decision that the AI system may influence. Transparency, fairness, accountability, explainability, privacy and data quality become operating requirements within that process. The implementation task is to connect them with existing model, conduct, data, outsourcing and operational-risk arrangements, rather than create an AI policy that sits beside the institution's responsibilities.

Build an inventory around real uses and decisions

The Guidance calls for an inventory of AI models, systems or technologies with material metadata including name, purpose and risk rating. Third-party models belong in that view. A useful inventory must show both the AI system and its accountability. The required fields include the business owner, consumer process, decision affected, data, model or provider, operating status, human role, validation evidence, controls and monitoring.

The inventory must distinguish the technology from each use. One platform may support internal research, communicate with customers and contribute to a high-impact decision. These uses create different consequences and therefore require different approvals, tests and monitoring. Linking them with existing model, outsourcing, data, conduct and operational-risk records prevents several registers from giving senior management conflicting views of the same exposure.

Make governance operational

The Guidance expects documented governance proportionate to the size, nature and complexity of the institution. Senior management and the Board are described as responsible and accountable for AI and ML systems and outcomes, with regular reporting on performance and risk. It also states that institutions should not employ AI models over which they have no control.

Governance becomes operational only when named people can make and challenge lifecycle decisions. The institution should assign authority to propose and classify a use, approve data and providers, design the evaluation, accept residual risk, release the system, approve material changes, respond to incidents and cease operation. Control functions need enough knowledge and evidence to challenge these decisions, while the business remains accountable for the consumer and operating outcome.

Identify high-impact consumer decisions

The Guidance defines a high-impact decision as a determination using AI that materially affects a customer's access to financial products or services, with a loan application and insurance claim given as examples. Institutions should trace where AI informs, recommends or makes such decisions and what happens when the output is wrong or disputed.

Risk rating should consider data quality and sensitivity, system capability, controls, impact and dependence on AI or third parties. The result should determine the approval route, independent challenge, testing, disclosure, human authority and monitoring required. A broad model category or vendor assurance report cannot replace the use-specific assessment.

Test fairness in the population and process

The Guidance expects AI and ML not to produce discriminatory or manipulative outcomes and says training data should be sufficiently accurate, relevant and representative of the customer population to which the model is applied. It also addresses periodic testing, including when a model is upgraded, materially changed or newly introduced.

A fairness review should define the consumer outcome, relevant groups, decision thresholds, error costs and process surrounding the model. It should examine representation, performance differences, approval or service outcomes, overrides and downstream effects. An overall accuracy figure can conceal a material disparity. However, a statistical difference does not identify its cause. The institution must examine the data, decision rule, surrounding process and plausible explanations before it selects a remedy.

Design transparency for the consumer decision

The Guidance addresses transparency where consumers interact with AI and particularly for high-impact decisions. It calls for understandable, accurate plain-language disclosures in Arabic and English and says institutions should consider measures to check understandability. It also raises opt-out rights for consideration, especially for high-impact decisions, in light of risk, fairness and feasibility.

The institution needs more than a general notice that AI is used. It should determine what the customer needs to know at each stage, how the decision can be explained accurately, how questions or disputes are handled and whether a different route is available. The explanation must reflect the real system and decision. One technical feature-importance method is not a complete consumer explanation.

Connect data quality, privacy, security and resilience

The Guidance sets expectations for data quality, assurance and operational resilience. The requirements cover accurate and current data, provenance and audit trails, personal-data protection, privacy and security by design, stress testing, validation, contingency planning and incident response. These concerns meet in the workflow. Data may be suitable for one purpose but not another, and a technically accurate model can still create consumer harm through unsuitable access, stale information or a failed hand-off.

The review should map data from collection and transformation through model use, output, action and retention. Qualified privacy, cybersecurity, legal and regulatory specialists should address their respective requirements. Model evaluation should use operating scenarios, including degraded inputs, unavailable services, attempted misuse and the fallback route.

Make human oversight meaningful

The Guidance distinguishes human-in-the-loop, human-on-the-loop and human-out-of-the-loop arrangements. It says the level of involvement should be commensurate with consumer risk and describes human-out-of-the-loop use as appropriate only for low-risk, non-material processes with controls.

The institution should test whether the person has time, information, competence, authority and a usable path to challenge or stop the output. An approval click does not create effective oversight when staff cannot understand the basis, are penalised for overriding or receive the case too late. Overrides, disagreements, escalations and consumer appeals should be recorded as operating evidence.

Control providers and automatic change

Outsourcing an AI system does not transfer the institution's responsibility for its consumer outcome. The Guidance therefore establishes expectations for third-party providers. These requirements address provider due diligence, access to relevant information, audit rights, data protection, cybersecurity, performance, termination, pre-deployment checks and third-party models in the inventory. It also expects institutions to understand and test automatic updates before implementation.

Procurement should establish the evidence needed throughout operation, not only at selection. Contracts and technical arrangements need to support notification of material changes, version visibility, evaluation, incident investigation, data and log access, concentration management and exit. A provider's general certification or benchmark does not establish fitness for the institution's consumer use.

Monitor outcomes and retain the ability to stop

The Guidance calls for continuing monitoring of reliability, relevance and consumer-protection outcomes, taking account of changes in data, markets and customer behaviour. It also addresses mechanisms to detect, report and remediate performance issues, bias and unintended consequences and the clear, immediate ability to cease use through human intervention.

Monitoring should connect technical behaviour to the customer and business process. Measures can include error types, subgroup outcomes, overrides, complaints, appeals, service completion, provider changes, drift, incidents and control failures. Thresholds should state who investigates, who can restrict or stop the system and what evidence is required before use resumes.

How Marketways supports an implementation review

Marketways can translate the Guidance into a use inventory, responsibility map, risk-classification method, control and evidence matrix, evaluation plan, monitoring design, vendor-review requirements and prioritised remediation programme. We connect the implementation programme to the institution's existing model-risk, consumer-protection, conduct, data, outsourcing, operational-risk and internal-assurance arrangements.

AI Governance and Model Risk provides the underlying service. AI Evaluation and Assurance tests system behaviour and controls for a defined use. AI Inventory and Shadow AI Audit helps establish coverage, while AI Impact Assessment and Risk Classification provides the general assessment method. Marketways does not provide legal or regulatory advice, certify compliance or replace the institution's authorised control functions and qualified advisers.

Evidence required for a CBUAE guidance review

An implementation review must compare documented expectations with the institution's actual controls and outcomes. The required evidence includes AI and model inventories, consumer-process maps, materiality methods, governance terms, validation and monitoring standards, provider contracts, privacy records, disclosures, human-review procedures, complaint and appeal data, incident records, internal audit findings and the institution's interpretation of applicable CBUAE requirements.

Independence and scope

Marketways independently advises banks and insurers on management, analytics and AI implementation. We use the published CBUAE material to help institutions examine workflows, controls, evaluation evidence and governance; we do not represent the Central Bank or issue a regulatory conclusion. Each institution must confirm its obligations with the responsible authority and qualified legal and compliance advisers.

Official sources

How Marketways can help

We put proportionate controls, monitoring and evidence around AI so management, auditors and regulators can rely on it.

Book a free 20-minute consultation →

References

  1. CBUAE Guidance Note on AI and ML
  2. CBUAE Governance and Accountability
  3. CBUAE Transparency, Data, Fairness and Security sections
  4. CBUAE Human Oversight and Consumer Protection
  5. CBUAE Continuous Monitoring and Review
  6. CBUAE Integration with Existing Frameworks
  7. CBUAE Model Management Standards, Model Governance