You are here:

CBUAE AI Governance Guidance: A CISO’s Guide to What Changes in 2026

Laptop displaying an AI governance dashboard, illustrating AI risk monitoring and oversight for UAE banks.

Ask a UAE bank whether it has an AI policy and the answer is almost always yes. Ask the same institution to name every AI system that touches a customer, explain who governs each one, and demonstrate that it could switch any of them off today, and the conversation slows down considerably.

That gap is what the Central Bank of the UAE’s Guidance Note addresses. It pulls cybersecurity, operational resilience and consumer protection into a single frame, and in doing so it hands security leaders a set of problems that conventional security controls were never built to solve. Access control and encryption do not tell you whether a credit model is disadvantaging a customer segment. Vulnerability management does not tell you whether a vendor silently updated the model behind your chatbot last Tuesday.

This blog translates the note into decisions a CISO can act on: what to inventory, what to test, what to demand from vendors, what to monitor, and what to put in front of the board. It also marks the boundary where the CISO’s remit ends, because accepting ownership of every AI risk in the institution is the fastest way to be accountable for outcomes you cannot control.

Is the CBUAE AI Guidance Note Mandatory?

No. The note is non-binding guidance written to help institutions formulate their own policies and procedures, and it supplements applicable laws, regulations and directives rather than replacing them.

Treating that as permission to wait would be a mistake. Three things complicate the picture. Much of what the note describes restates obligations that already bite under separate instruments, including the CBUAE Model Management Standards and the Consumer Protection Regulation. Supervisors now have a published articulation of what good practice looks like, and supervisory dialogue tends to follow published expectations. And the direction of travel across the UAE is one way only, with DIFC Regulation 10 already fully enforced and a new Federal Authority for AI and Data holding a mandate to propose legislation.

Read the note alongside your existing obligations. Keep a clear line in your own documentation between what the note recommends and what separate law already requires, because those two things attract different consequences when you fall short.

Separate What You Own From What You Are Being Handed

The usual first reaction is that AI governance belongs to model risk or compliance. That reading does not survive contact with the text, which ties security and privacy safeguards to robustness, stress testing, incident response, operational resilience and oversight of outsourced AI, then places AI risk inside enterprise-wide risk management alongside cybersecurity, operational, credit and conduct risk.

Three distinct contributions follow, and keeping them separate protects you:

  • Security work proper. Protect the data, systems and access paths AI depends on. Assess exposure. Support resilience and response when something breaks.
  • Shared governance work. Help define what evidence a deployment needs, what controls apply to third parties, what monitoring runs, and how escalation works.
  • Work that belongs elsewhere. Business suitability, fairness, customer disclosures, complaints and redress need people outside security at the table.

The note assigns accountability to the board and senior management, and asks for defined roles across IT, risk management, internal audit and the relevant committees. It describes no CISO-only model. Help make AI governance operational. Decline to become its default owner.

Build an Inventory That Supports a Decision, Not a Count

The note asks for an inventory of AI models, systems and technologies, including third-party models, recording at minimum the model name, purpose and risk rating. Risk assessment is expected to weigh data sensitivity, the capability of the AI, the controls around it, its impact and the dependency it creates.

That is the floor, not the target. Find the AI built in-house, then go looking for the AI that arrived inside software you already bought. The second group is usually larger and almost always less documented, because nobody treats a feature release note as a procurement event. Tie each deployment to a business purpose and a named owner. Flag the customer-facing uses and the high-impact decisions. Where one model serves several applications, record that relationship, because a fault in the model becomes a fault in all of them simultaneously.

The fields below go past the minimum and are the ones that earn their place in practice:

FieldThe question it answers
Business and technical ownersWho is accountable for this use case and for running it?
Customer impactCould this materially affect access to a product or service?
Data categories and flowsWhat information goes in, and where does it travel?
Vendor and dependenciesWhich third parties sit underneath this?
Human oversight arrangementWho can review or override an outcome?
Assurance and review statusWhat evidence supports continued use?
Stop and fallback arrangementsHow do we cease use without unmanaged disruption?

An inventory that cannot answer those seven questions is a list of tools. It will not survive a supervisory conversation.

Scale Assurance to Harm, Not to Budget Cycles

The note calls for security and privacy by design, stress testing and validation, safeguards against unauthorised access or misuse, and pre-deployment checks where third parties are involved. It addresses testing for bias annually, and again when a model is upgraded, materially changed or newly introduced.

The temptation is to commission one generic AI security test and file the report. That approach fails the first time anyone asks what the test actually covered for a specific use case. Start with the system’s purpose, its data and the harm it could cause, then pick the assurance activity.

Assurance areaThe question to answer
Data protectionCan sensitive information be accessed or disclosed inappropriately?
Access and authorityCan the system act beyond the permissions it was granted?
RobustnessHow does it behave under unusual, misleading or adversarial input?
Consumer outcomesCould a wrong output mislead or disadvantage a customer?
ResilienceWhat happens when the model, the integration or the provider fails?

For generative AI applications, testing may extend to prompt manipulation, disclosure of sensitive data and misuse of connected tools. Those are implementation choices serving the broader security and robustness objectives in the note. CBUAE names no specific attack techniques, and telling your board otherwise will not hold up when someone reads the source.

Treat Vendor AI as Retained Risk

Institutions remain responsible for AI functions they outsource. The note addresses provider due diligence, information and audit rights, notification of material developments, termination provisions, pre-deployment checks and standards for third-party models. It also describes annual cybersecurity reviews by independent, suitably qualified third parties in the procurement and deployment context.

Five questions expose whether oversight is real or decorative:

  • Can we obtain enough evidence to assess the service as deployed, rather than as marketed?
  • What happens when the provider changes the model or its functionality without telling us?
  • Could this vendor actually support an incident investigation on our timeline?
  • Are our audit and information rights usable in practice, or only written down?
  • Can we discontinue the service and keep operating?

The question is never simply whether the vendor is secure. It is whether the institution can get sufficient assurance and keep meaningful oversight of something it does not control. Get the answers in writing before signature, because your negotiating position disappears afterwards.

Make Intervention Real Before You Need It

This is the capability institutions most often cannot demonstrate, and it is the one a supervisor can test in a single question.

The note asks for continuous monitoring, mechanisms to detect and remediate performance problems, bias and unintended consequences, and a clear and immediate human ability to cease AI use, with human oversight proportionate to consumer risk.

Making that real takes more than a policy paragraph. Set thresholds that trigger investigation, restriction and suspension. Name who holds authority to stop a system, and confirm that person is reachable outside business hours. Test whether disabling the AI also halts the downstream actions it already set in motion, because a stopped model and a stopped process are different things. Define the fallback. Connect your security incident process to model failures and customer-harm escalation, which today usually run in separate channels. Route customer challenges to a human review that can actually change the answer.

Consider a hypothetical. An AI-enabled customer service application starts giving misleading product information after a provider update nobody flagged. The response involves restricting the application, preserving evidence before it rotates out of logs, bringing in business and compliance colleagues, and working out whether affected customers need contacting and putting right. Each step has an owner. Most institutions cannot name all of them today.

Report Evidence, Not Maturity Scores

The note asks for regular performance and risk reporting to senior management and the board, with AI risk integrated into existing governance and risk appetite. It prescribes no dashboard and no reporting frequency, which leaves room to report something genuinely useful.

What tends to be worth putting in front of a board:

  • Material AI use cases and who owns them
  • High-impact uses where assurance is complete, and where it is not
  • Outstanding findings and any accepted exceptions
  • Significant provider changes still awaiting review
  • AI incidents and customer-impact issues
  • Monitoring breaches and the state of remediation
  • Stop and fallback arrangements that have been tested, with the date of the test

That last line carries more weight with a board than any maturity rating, because it is the only item on the list that cannot be produced from a slide template.

Sequence the Work Proportionately

The note expects governance proportionate to the institution’s size and complexity, so sequencing is yours to set. The phasing below is illustrative rather than a CBUAE timetable.

PhaseFocusOutputs
Establish the baselineIdentify AI use, owners, customer impacts and existing controlsInitial inventory, responsibility map, gap assessment against the note
Prioritise material exposureReview high-impact uses, sensitive data flows and critical vendorsPrioritised assurance plan, vendor evidence gaps, deployment criteria
Operationalise oversightStand up monitoring, escalation, intervention and reportingMonitoring procedures, tested stop and fallback arrangements, board reporting

Summary

The CBUAE Guidance Note does not create a new compliance regime, and treating it as one misreads both its status and its usefulness. It describes what responsible AI looks like in a regulated financial institution, and it does so in terms a security leader can act on: know what AI you run, scale assurance to the harm a system could cause, keep real oversight of vendor AI, monitor in production, and retain the ability to stop. The CISO contributes the security and assurance foundations. The board and senior management stay accountable for the outcomes. Institutions that already run those foundations will absorb whatever the Federal Authority for AI and Data issues next. The rest will build them under a deadline.

Frequently Asked Questions

Is the CBUAE AI Guidance Note legally binding?

No. It is non-binding guidance that supplements applicable laws, regulations and directives. In practice, expect supervisors to assess you against it, and note that much of its content restates obligations that are binding under separate instruments.


Who does it apply to?

All licensed financial institutions within the CBUAE supervisory ecosystem, including insurance providers. It covers AI, machine learning and generative AI, not only customer-facing chatbots.


Does this make the CISO responsible for AI governance?

No. The note assigns accountability to the board and senior management, and asks for defined roles across IT, risk management, internal audit and relevant committees. The CISO contributes security and assurance capability inside that structure.


Do all AI decisions need human approval?

No. Human oversight is expected to be proportionate to consumer risk. Operating without a human in the loop is described for low-risk, non-material processes with appropriate controls, which means the design question is about materiality rather than a blanket rule.


Does CBUAE require prompt-injection testing or annual red teaming of every model?

No. The note addresses annual bias testing and annual independent cybersecurity reviews in their own contexts. Prompt-injection testing is an implementation choice that can serve the broader security and robustness objectives, not a named requirement.


If the model belongs to our cloud provider, is the risk theirs?

No. Institutions remain responsible for outsourced AI functions. The relevant question is whether you can obtain enough evidence to assess the deployed service and retain practical oversight, including audit rights you could actually exercise.


Do customers have an absolute right to opt out of AI decisions?

Not absolute. The note asks institutions to consider opt-out rights with regard to risk, fairness and feasibility, and separately addresses human review and alternative arrangements. Treat it as a design consideration per use case rather than a universal entitlement.


Where should an institution with no AI inventory start?

With the inventory, scoped first to customer-facing systems and high-impact decisions. Assurance, monitoring and intervention all depend on knowing what exists, and a partial inventory covering your riskiest systems is more useful than a complete one that arrives in six months.


How does this interact with DIFC Regulation 10 or ISO/IEC 42001?

A DIFC-registered entity answers to Regulation 10 regardless of this note. ISO/IEC 42001 is voluntary and gives you one management system producing evidence that serves several regimes at once, which is usually cheaper than running parallel compliance programmes.

Author

Seecko Das

Seecko Das is an information security, Governance, Risk, and Compliance consultant with a proven record of securing critical infrastructures and enabling regulatory confidence across the MENA, EU, and Asian regions. He specializes in advising fintech, healthcare, cloud, commercial gaming, and high-data-value organizations on aligning technology operations with international security, privacy, and AI governance standards. He holds certifications in ISO 27001/42001 Lead Auditor, CISA, PCI QSA, PCI SSLCA, and CEH, and brings deep expertise across audit, governance, and assurance disciplines. His experience spans PCI DSS/3DS/PIN and SWIFT CSP certification programs, ISO 27001/27701/42001 implementations, EU AI Act and NIST AI RMF adoption, WLA SCS audits, and compliance with UAE IAR, DESC ISR, GDPR, UAE PDPL, and DPDPA requirements. Seecko combines technical rigor with strategic oversight to help organizations manage emerging AI and cyber risks while achieving sustainable compliance and market trust.

Table of Contents

Protect Your Business from Cyber Threats Today!

Safeguard your business with tailored cybersecurity solutions. Contact us now for a free consultation and ensure a secure digital future!

Ready to Secure Your Future?

We partner with ambitious leaders who shape the future, not just react to it. Let’s achieve extraordinary outcomes together.

I want to talk to your experts in:

Related Blogs

Business leaders reviewing a holographic AI display with data and lock icons, illustrating ISO 42001 compliance and risk management.
Silhouettes interacting with an AI network graphic, illustrating AICPA's 2026 SOC 2 guidance for AI systems.