AI governance for Qatar financial institutions: working to the QCB AI Guideline

The Qatar Central Bank’s Artificial Intelligence Guideline entered into force on 4 September 2024 for every entity QCB licenses. It is more prescriptive than most financial sector AI guidance. QCB wants to see the register, approve high-risk systems before they go live, and know that a trained person can stop any of them. This guide sets out an order of work for a bank, insurer or other licensed entity that wants to get there.

Start with a complete AI register

Almost everything in the Guideline hangs off the register in section 10. Classification needs it. Approvals are triggered by what is on it. And the full register goes to QCB every year and on request.

The hard part is completeness. Entities usually know about the credit model and the fraud engine. They know less about AI inside vendor products, staff use of public generative AI tools and features switched on in software already licensed. Those count too, because the Guideline covers AI an entity purchases and processes it outsources that rely directly on AI. Our guide on finding shadow AI covers where to look.

For each system, the Guideline asks you to record four things.

• Whether it is high-risk or not.
• Whether you are a user or a provider of the system.
• A category reflecting its function.
• The human oversight protocol used.

For anything bought, licensed or outsourced, add the provider and any outsourcer, a third-party supplier assessment, the provider’s country of registration and registration number, governing law, creation and upgrade dates, and contract start and renewal dates. High-risk systems also need an assessment of how easily the AI could be substituted, an alternative provider where one exists, a detailed life cycle description, and the date and results of the latest risk assessment or audit. Our guide to building an AI system register covers the general method.

Classify every system

Section 9 asks entities to assess the risk and criticality of the processes each AI system supports, and to set an AI risk score. Two rules decide most cases.

First, autonomy raises risk. A system that acts without direct human oversight will probably be judged higher risk than one that produces AI-assisted outcomes for a person to decide on. Second, three uses are high-risk whatever the score. These are decisions on consumer access to financial services, internal decisions that materially affect employees, and processing of sensitive personal information. Credit scoring, onboarding checks, claims triage, CV screening and performance analytics will usually land here.

Write your classification criteria down. Section 10 requires you to disclose them to QCB, along with a high-level risk and impact assessment, so they need to be consistent and defensible.

Plan for QCB approval

Section 11 sets three approval triggers. You need QCB approval before launching a new AI system as a provider, before any material modification of one, and before signing any purchase, licensing or outsourcing agreement for a high-risk AI system. For new high-risk systems, section 14 adds that the full results of training, validation and testing go to QCB, and a user must also submit the provider’s instructions, its technical and training results, its own testing results and its human oversight plan.

That has two consequences for project planning. Approval has to sit on the critical path of any AI project, with time allowed for QCB to send a system to its sandbox for further evaluation. And procurement needs a hold point, so that a high-risk AI contract is not signed before approval arrives. Section 12 also requires board approval for outsourcing any AI function, due diligence on the provider, and service agreements with audit, confidentiality, continuity and exit terms. Our vendor AI risk guide sets out how to run that due diligence.

Give every system an oversight protocol

Section 13 requires a human oversight protocol for every AI system, chosen from three. With AI-assisted decision-making, a person approves every designated output. With human exception monitoring, a person monitors and can intervene or shut the system down. A fully autonomous system runs without a person in the loop.

For high-risk systems, the named supervisor must be trained, understand the system’s limits, be able to interpret its output, decide to disregard or reverse it, and stop it. A fully autonomous high-risk system needs QCB approval before launch, built-in guardrails it cannot override, set limits linked to warning levels or auto-close routines, and a supervisor who can shut it down when outputs look wrong. Our guide to human oversight of AI covers how to make that oversight real rather than nominal.

Prepare customer disclosure and recourse

Part E is where customers see the Guideline. Entities must tell customers when they are interacting with an AI system, describe AI use in plain language, and explain on request the types of data and factors that drive an AI decision. They must also say how a decision may affect the customer and whether it can be reversed. Customers must be able to raise questions about AI decisions and ask for a review of decisions made without human involvement, either by resubmitting corrected data or by asking a qualified person to review a negative decision. Where you choose not to offer an opt-out, another route to review is required.

These pieces cut across product, legal, complaints and operations teams, so assign an owner for each customer-facing AI system early.

Keep the record current

The Guideline expects ongoing work, not a one-off project. Section 19 asks for audit logs, design documentation, model version records and archived training data, and a serious incident must go to QCB immediately once a causal link to the AI system is established or reasonably likely. Exemptions must be requested from QCB, recorded in an exemptions register and given an expiry date. Board and senior management stay accountable throughout, so a regular AI report to the board, built from the same register, closes the loop.

Personal data used by any of these systems is also governed by Qatar’s PDPPL, which the Guideline names. Under the NCSA’s guidelines, automated decisions that limit access to a product or service are a sign that a data privacy impact assessment is needed.

Where Aicura fits

Aicura’s AI register holds each system with its owner, provider, risk rating, oversight setting and status, and discovery surfaces AI that is not yet on it. Risk and impact assessments, vendor records, guardrails and monitoring sit on the same records, so the information QCB asks entities to keep is in one place. Aicura supports the work the Guideline describes. It does not assess against the Guideline or file with QCB, and the decisions stay with you.