CBUAE’s AI Guidance Sends a Clear Message to GCC Banks

A customer opens your bank’s app and asks the AI assistant a question. The AI gives them the wrong answer. They act on it and lose money.

Who pays?

On 23 February 2026, the Central Bank of the UAE made that accountability expectation explicit for the licensed financial institutions it supervises. Its new Guidance Note on AI and machine learning is non-binding, but the message is clear: AI risk now sits with the institution’s own governance structure, including the board and senior management.

The regional pattern is consistent. The Qatar Central Bank’s 2024 AI Guideline points to the same supervisory logic. Saudi Arabia has not yet issued an equivalent AI-specific rulebook for financial institutions, but SAMA’s existing frameworks already create relevant expectations around governance, outsourcing, cyber, data, and model oversight.

The Air Canada warning

In November 2022, a customer used Air Canada’s website chatbot to ask about bereavement fares. The chatbot told him he could apply retroactively. He booked, paid full fare, then applied. Air Canada refused, citing its actual policy. The customer brought the dispute to the British Columbia Civil Resolution Tribunal, which issued its decision in 2024 (Moffatt v. Air Canada, February 2024).

Air Canada argued the chatbot was “a separate legal entity that is responsible for its own actions.” The Tribunal rejected the argument: the airline was responsible for the information on its website, whether it came from a static page or a chatbot.

The damages were small, but the signal was hard to miss. The case is not binding in the GCC, and it does not determine how regional courts or regulators would treat a similar dispute. But it describes the problem financial regulators are now turning into supervisory expectation: an institution cannot put a model in front of a customer and then disown the answer.

Three principles in the CBUAE Note

The CBUAE Guidance Note turns that principle into a practical supervisory expectation for financial institutions, and adds two more.

First, accountability is allocated to boards and senior management. The guidance expects institutions to take ownership of AI outcomes, governance structures, and alignment with risk appetite. AI risk is enterprise risk, not a technology line item. The same control functions that own credit, market, compliance, operational, cyber, and technology risk are expected to be capable of challenging AI-driven processes.

Second, the guidance introduces the concept of a “high-impact decision”: any AI-driven determination that materially affects a customer’s access to financial products or services. Credit approvals. Pricing. Insurance claims outcomes. Fully automated credit or insurance decisions without the possibility of human intervention, the guidance warns, are unlikely to meet supervisory expectations.

Third, the outsourcing principle is stated plainly: regulatory responsibility cannot be outsourced. Institutions remain accountable for third-party AI systems, must secure audit rights contractually, and must retain the ability to suspend or terminate systems if required. The supervisory expectation is shifting from live monitoring alone to after-the-fact accountability: can the institution explain, reproduce, and audit the AI-driven outcome when challenged?

The regional picture: this is not just the CBUAE

The Qatar Central Bank’s AI Guideline, issued in 2024, already requires QCB-licensed financial institutions to maintain an AI systems registry, obtain prior QCB approval before launching new AI systems, and notify customers when they are interacting with AI. The registry expectation matters because it turns AI governance into an operating discipline: knowing what AI you have, where it runs, who owns it, what data it touches, and what risk category it falls into.

Saudi Arabia is approaching the issue through a different regulatory route. SAMA has not yet issued a standalone AI rulebook for financial institutions. But its existing rulebook already creates relevant obligations around cybersecurity governance, third-party risk, material outsourcing approval, cloud approval, data-location controls, audit and review rights, termination rights, and model oversight.

Saudi Arabia’s broader AI governance also sits within the Personal Data Protection Law, SDAIA guidance, and sector-specific supervision. For banks, the practical implication is clear: AI systems will not be judged only as technology deployments. They will be judged through the existing lenses of customer data, outsourcing, operational resilience, credit models, cyber risk, and accountable governance.

What a CIO should be asking this quarter

The implication is architectural, not just procedural. Supervisors are increasingly interested in how AI behavior is constrained at the point of execution, not just how the model was trained.

Can the bank produce, today, a complete inventory of every AI system in production or pilot, including vendor, owner, data flows, risk classification, and fallback arrangement? If the answer is “not in one place,” that gap is likely to become one of the first supervisory findings.

Which AI-driven workflows currently make or influence “high-impact decisions” as the CBUAE defines them, and does each one have meaningful human review at the point of decision? The CBUAE distinguishes between human-in-the-loop, human-on-the-loop, and fully autonomous systems, and strongly signals a preference for meaningful human oversight where customer access to a financial product is at stake.

Do the existing contracts with AI and cloud vendors give the bank the audit, inspection, notification, and termination rights that the new guidance assumes? Many banks should assume the answer is no until legal, procurement, risk, and technology have reviewed the contracts together. Treating this as an IT review will miss the point; the real work usually sits across procurement, legal, risk, and technology.

Where this leaves the AI roadmap

The regulator is not saying AI cannot be used. The CBUAE frames the guidance as a balance between technological advancement, consumer protection, and financial stability. The QCB and SAMA positions are similar in posture, even if their frameworks differ in specificity.

The shape of the AI roadmap question has changed. It is no longer “where can we deploy AI?” It is “where can we deploy AI in a way we are willing to be accountable for at board level?”

Most banks do not fail AI governance because of models. They fail because ownership, inventory, procurement controls, and escalation paths are fragmented across the organization. The institutions that answer the new question well are not necessarily the ones with the most ambitious AI strategies. They are the ones that have already done the unglamorous work the new guidance assumes is already done.

The banks doing that work this quarter will be closer to production next year. The banks treating it as a compliance afterthought will still be running pilots.

The practical answer is architectural. Banks need AI systems whose behavior can be constrained before execution, reconstructed after execution, and defended under scrutiny.

That is the problem DataKite is built for: governed AI infrastructure for banks operating in regulated environments.