Birbal AI

Guide · AI in financial services

AI in UAE banking and finance: the Central Bank's rules and what to build

The Central Bank's guidance note of February 2026 sets out how banks, insurers, finance companies and other licensed financial institutions should govern, explain and oversee the AI they use with customers, and banks' decision models also fall under the Model Management Standards. This guide explains both: what the note asks, the three models of human oversight, what customers can ask for, how to treat AI from vendors, and the records and controls your systems need, from the Central Bank's rulebook, checked on 19 November 2026. It is general information, not legal advice.

Updated · 10 min read · By the Birbal AI team, Abu Dhabi · اقرأ بالعربية

Key takeaways

  • On 11 February 2026 the Central Bank issued a guidance note on consumer protection and the responsible adoption and use of AI and machine learning by licensed financial institutions, including insurance providers. It sets principles that institutions are expected to build into their own AI policies, on top of the laws and rules that already apply.
  • The board and senior management are accountable for AI systems and their outcomes, institutions should not employ AI models they have no control over, and every AI model, system or technology they develop or deploy goes into an inventory with at least its name, purpose and risk rating.
  • Customers should be told when AI is used, above all for high-impact decisions such as a loan application or an insurance claim, in plain Arabic and English. Deployed AI is tested for bias at least once a year and after material changes, and customers can ask for a human review or an explanation of an AI decision.
  • Banks' decision models, AI included, also follow the Model Management Standards: an inventory that covers third-party models, at least two tiers of model risk, and regular independent validation, which must be done by a party other than the provider when a third party supplied the model's methodology.

What the guidance note is

The Central Bank issued the Guidance Note on the Consumer Protection and Responsible Adoption and Use of Artificial Intelligence and Machine Learning by Licensed Financial Institutions in the U.A.E. on 11 February 2026 and announced it on 23 February 2026. It is addressed to licensed financial institutions, including insurance providers, and focuses on the areas where AI may affect consumers, with the aim of promoting consumer protection and good market conduct.

Its definition of AI covers systems and models that perform tasks or generate outputs that normally require human intelligence, and includes generative AI, the models that understand and generate human-like text, audio and images. It defines a high-impact decision as any determination by an institution using AI that materially affects a customer's access to financial products or services, for example a loan application or an insurance claim.

The note sets principles and guidelines rather than new rules. The Central Bank expects institutions to refer to it as they develop their own policies on the ethical and responsible use of AI with consumers, and the note supplements, rather than replaces, the laws, regulations and directives that already apply. It has ten sections; sections 2 to 9 carry the expectations:

Guidance note of 11 February 2026, CBUAE Rulebook.
SectionWhat it asks
2. Governance and accountabilityA documented AI governance framework proportionate to the institution; the board and senior management accountable for AI systems and outcomes; no AI models the institution has no control over; control functions able to challenge AI outcomes; an inventory of all AI models, systems and technologies.
3. Fairness, non-discrimination and ethicsNo discriminatory or manipulative outcomes; representative training data; bias tests once a year and on every upgrade, material change or new model.
4. Transparency and explainabilityOpenness about the use of AI, particularly for high-impact decisions; plain-language disclosures in Arabic and English; opt-out options considered.
5. Data quality, privacy and securityAccurate, relevant and up-to-date data with clear provenance and audit trails; privacy and security by design.
6. Continuous monitoring and reviewModels monitored, updated or stopped as conditions change; the clear and immediate ability, with human intervention, to stop any AI system; responsibility kept for outsourced AI.
7. Human oversight and consumer protectionMeaningful human oversight under one of three models; human review or an explanation on request; alternatives to AI decision-making; complaints and redress.
8. Integration with existing frameworksAI risk inside enterprise risk management and the conduct risk framework, with reporting to the board and regulators; a risk rating for each AI system.
9. Outsourcing and third-party riskDue diligence and contracts for AI vendors and cloud providers, checks before deployment, and the same standards as for in-house models.

Governance, accountability and the AI inventory

  • A documented AI governance framework, proportionate to the institution's size, nature and complexity, and a culture of responsible use of AI that understands its risks.
  • The board and senior management are responsible and accountable for AI systems and their outcomes, including model selection, deployment, resourcing and ongoing oversight. Institutions should not employ AI models they have no control over.
  • Risk committees and control functions such as compliance, internal audit and risk management understand AI-driven processes and can challenge their outcomes, and the adoption and use of AI is part of the risk management framework.
  • An inventory of all AI models, systems and technologies the institution develops or deploys, with all material metadata, at least the model name, purpose and risk rating, kept in line with the Model Management Standards and Model Management Guidance.
  • AI risk sits inside enterprise-wide risk management: consumer risk from AI-driven models is treated as part of the conduct risk framework, with reporting to the board and regulators, and each AI system or application gets a risk rating.

Fairness, transparency and what customers can ask for

  • No discrimination or manipulation: AI should not produce discriminatory or manipulative outcomes against individuals or groups, and no AI system should be deployed or used if it is discriminatory or manipulative, or becomes so after deployment.
  • Representative data: data used to train models should be accurate, relevant and representative of the customers the models will be applied to, with clear provenance and audit trails, and privacy and security built in by design.
  • Bias testing: deployed AI is tested for embedded biases and discriminatory outcomes once a year, and each time a model is upgraded, materially changed or a new one is introduced.
  • Transparency: institutions are open about their use of AI, particularly for high-impact decisions, tell customers when they are communicating with an AI application, and are clear about how their AI systems operate and make decisions.
  • Language: disclosures are in understandable plain language, in both Arabic and English, with telephone support in all major languages of the UAE.
  • Opt-out: institutions should consider letting customers opt out of AI, particularly for high-impact decisions, taking into account the risk to the customer, fairness and feasibility.
  • Human review: customers can ask for a human review or an explanation of an AI-generated decision, an alternative exists for customers who do not wish to be subject to AI decision-making, and complaints and redress have clear channels.

Human oversight: three models

Section 7 asks that AI systems operate under meaningful human oversight and judgement, particularly for decisions with significant implications for consumers, and describes three models:

Guidance note of 11 February 2026, section 7.
ModelHow it works
Human-in-the-loopAI provides recommendations, and a person keeps full authority to approve or reject the outcome.
Human-on-the-loopAI works on routine tasks by itself, while people monitor the outcomes and can step in.
Human-out-of-the-loopAI operates without direct human involvement, which the note limits to low-risk, non-material processes with appropriate controls in place.

Whatever the model, the note asks institutions to monitor and review their AI as data, market conditions and customer behaviour change, to update or stop using a model where needed, to have ways to detect, report and fix performance issues, bias and unintended consequences, and to keep the clear and immediate ability, with human intervention, to stop using any AI model, system, technology or application.

AI from vendors and cloud providers

Institutions remain responsible for outsourced AI functions. Where they rely on third-party vendors or cloud service providers for AI and machine learning, the note asks for:

  • Due diligence on the provider's reputation in AI and on its governance, security and data protection practices.
  • Contracts that give the institution audit and information rights, with terms on termination, data protection, cyber security and performance.
  • Checks before a vendor's AI is deployed, and independent cyber security reviews of the provider.
  • The same standards of fairness, explainability and robustness for models developed or hosted by third parties as for models built in-house.

For banks, the Model Management Standards add their own rules: the model inventory covers models provided by third parties, third-party consultants support the steps of a model's life-cycle but are never responsible for them, and when a third party provides the methodology used to develop a model, any later validation must be done by a different party. Customer data rules also shape where AI can run: the Consumer Protection Standards require consumer and transaction data to be held and stored in the UAE.

The Model Management Standards for banks

The Central Bank issued the Model Management Standards and the Model Management Guidance in late 2022. The standards set the modelling practices that banks in the UAE must implement if they decide to use models for decision-making, and the guidance note points to them for the AI inventory and for the governance, use and validation of AI. They are built around seven components: model governance, data management, model development, model implementation, model usage, performance monitoring and independent validation.

  • Scope: model governance covers all models used to make decisions within the bank, with dedicated documentation for each step of a model's life-cycle.
  • Tiers: models are grouped by model risk into at least two tiers, Tier 1 for the more critical models and Tier 2, and IFRS 9 models for large portfolios and capital forecasting models must be Tier 1.
  • Inventory: a comprehensive inventory of all models used in production to support decision-making, covering internal models and models provided by third parties.
  • Independent validation: every model is validated independently on a regular basis, by an internal team or a third party, kept separate from development, and the validation team reports its findings to senior management and the board every year.

The standards are written for banks. The guidance note asks every licensed institution, insurers and finance companies included, to keep its AI inventory in line with them.

What to build: the records and controls

Most of the note's expectations end up as features of the systems that use AI. A bank, insurer or finance company putting AI in front of customers typically needs:

  • An AI register: every model, agent and vendor AI service with its owner, version, purpose, data sources and risk rating, including generative AI assistants used by staff.
  • Decision records: for each AI-supported decision, the inputs, the model version, the score or recommendation, the reasons, and who approved or overrode it, so it can be explained, reviewed and audited later.
  • Approval queues for human-in-the-loop decisions, and alerts, sampling and dashboards for processes that run with people on the loop.
  • Explanations and disclosures in Arabic and English: reasons staff can give a customer, and messages that say when a customer is dealing with an AI application.
  • A route to a person: human review on request, an alternative without AI, and complaints linked to the decision they concern.
  • Tests before release and on a schedule: accuracy, bias and robustness checks before go-live, once a year and after every material change, with results kept for validation and audit.
  • A stop switch: an authorised person can stop any model or AI application at once, and the process falls back to people.
  • Models under your control: open-weight models on your own servers or in a UAE data center, or vendor services under contracts with audit and information rights.

Birbal AI builds these controls into the systems we deliver for UAE banks, insurers and finance companies. Warsha's AI agent governance gives every agent an owner, a version, tests before release, human approval of high-impact actions and a trace of every run, on your own servers; Mustashar answers staff from your own documents, with sources, on servers inside your institution; our credit decision engines record the reasons for every decision; and Nida says it is an AI assistant and hands customers over to your team. We are a technology provider, not a law firm or a model validator: your compliance team and legal adviser decide how the guidance applies to you, and independent validation stays with your validation team or a party you appoint.

Questions and answers

Does the Central Bank of the UAE regulate AI?

Yes. Its guidance note of 11 February 2026 sets out how licensed financial institutions, including insurers, should govern and use AI and machine learning where it may affect consumers: governance and accountability, fairness, transparency, data, monitoring, human oversight and third-party risk. Banks' decision models also fall under the Model Management Standards issued in 2022.

Is the CBUAE AI guidance note mandatory?

It sets principles and guidelines rather than new rules: the Central Bank expects institutions to refer to it when they write their own AI policies, and it supplements the laws, regulations and directives that already apply. The Model Management Standards are different: banks must implement them if they use models for decision-making.

What is a high-impact decision?

Any determination by a licensed institution using AI that materially affects a customer's access to financial products or services, for example a decision on a loan application or an insurance claim. The note's expectations on transparency and opt-out weigh most for these decisions.

What are human-in-the-loop and human-on-the-loop?

Two of the note's three models of human oversight. In human-in-the-loop, AI recommends and a person keeps full authority to approve or reject the outcome. In human-on-the-loop, AI works on routine tasks by itself while people monitor and can step in. AI without direct human involvement is limited to low-risk, non-material processes with appropriate controls.

How often should AI be tested for bias?

The note asks institutions to test deployed AI once a year, and each time a model is upgraded, materially changed or a new one is introduced, to find and correct embedded biases and discriminatory outcomes.

Can a UAE bank use a third-party or cloud AI model?

The note does not rule it out, but asks for due diligence on the provider, contracts with audit and information rights, checks before deployment and the same standards of fairness, explainability and robustness as for in-house models, and it says institutions should not employ AI models they have no control over. The Consumer Protection Standards also require consumer and transaction data to be held and stored in the UAE.

Does the guidance apply to insurers and finance companies?

Yes. It is addressed to licensed financial institutions, including insurance providers, so it covers banks, insurers, finance companies and the other institutions the Central Bank licenses. The Model Management Standards are written for banks, but the note asks every institution to keep its AI inventory in line with them.

Is this guide legal advice?

No. It is general information from the Central Bank's rulebook and its announcement of the guidance note, checked on 19 November 2026. Confirm how the guidance applies to you with your compliance team, your legal adviser and the Central Bank.

Sources

  1. Guidance Note on the Consumer Protection and Responsible Adoption and Use of Artificial Intelligence and Machine Learning by Licensed Financial Institutions in the U.A.E., 11 February 2026 (CBUAE Rulebook)
  2. Guidance Note on AI and machine learning, full text as PDF (CBUAE Rulebook)
  3. CBUAE issues guidance note to protect consumers and ensure responsible use of AI in the financial sector, 23 February 2026 (Central Bank of the UAE)
  4. Model Management Standards and Guidance: context and objective (CBUAE Rulebook)
  5. Model Management Standards, 1.1: regulatory context (CBUAE Rulebook)
  6. Model Management Standards, Part I: general standards (CBUAE Rulebook)
  7. Model Management Standards, 4: model governance, tiers and inventory (CBUAE Rulebook)
  8. Consumer Protection Standards, 6.1.6: retention and storage of consumer data (CBUAE Rulebook)

Birbal AI is an independent company based in Abu Dhabi. Facts about third parties come from the sources above; check them for the latest position.

Free briefing

Putting AI in front of customers? Book a free briefing.

A session with our team on the AI use case you want to run or fix first, what the guidance note and the Model Management Standards ask of it, and how we would build it inside your controls.

Book an AI briefing

We only use your details to reply to you. Arabic or English, as you prefer.