Regulation guide
QCB Artificial Intelligence Guideline: what it requires of AI agents at Qatari financial institutions
The Artificial Intelligence Guideline is the Qatar Central Bank's rulebook for how banks, insurers and other QCB regulated financial institutions build, buy, outsource and run AI. In force since 4 September 2024, it asks for a board owned AI strategy, a register of every AI system, a risk classification with a defined class of high risk AI, a human oversight protocol per system, and QCB approval before new AI systems go live.
Read the source text (Qatar Central Bank)Where it stands: In force since 4 September 2024 for every entity regulated by the Qatar Central Bank, with no amended version found as of 17 November 2026; in 2025 QCB added a Data Handling and Protection Regulation that requires its approval before data is processed in fully autonomous decision making systems.
- Issued by
- Qatar Central Bank (QCB)
- Type of instrument
- Supervisory guideline; its text calls its contents instructions
- Applies to
- Every entity regulated by QCB that develops, buys or outsources AI
- In force since
- 4 September 2024
- Enforced under
- QCB Law, Law No. (13) of 2012
- Maximum QCB fine
- QAR 10 million per violation, plus up to QAR 100,000 a day while it continues
Checked against the sources on . This guide explains the rule and is not legal advice.
Overview
For conversational and agentic AI it is direct. Customers must be told they are dealing with AI, can ask what drives a decision, and can have a negative AI decision reviewed by a qualified human. Companion QCB regulations keep personal and financial data in Qatar unless QCB approves otherwise. This page explains the rules and is not legal advice.
Timeline
January 2018
QCB issues its Technology Risks instructions for banks, which the AI guideline lists among the rules Entities must also follow. (AI Guideline, Section 24.1.3)
29 October 2019
The government adopts the National AI Strategy, a policy document.
February 2024
The National Cyber Security Agency (NCSA) publishes its voluntary Guidelines for Secure Adoption and Usage of AI, version 1.0. (NCSA Guidelines, Document Control)
15 April 2024
QCB's Cloud Computing Regulation enters into force. (Cloud Computing Regulation, Section 1)
4 September 2024
The AI Guideline enters into force; its text sets no transition period. (AI Guideline, Section 1)
Who and what the guideline covers
It applies to every Entity, meaning an entity regulated by QCB (Section 2, term 9), when it develops AI itself, buys AI systems, or outsources functions that rely directly on AI (Section 5). QCB groups the institutions it regulates as banks and insurers on one side and exchange houses, fintech, finance and investment companies and insurance brokers on the other (Data Handling and Protection Regulation, Appendix 1). AI covers technologies that simulate human traits such as reasoning and learning and produce an output or decision, such as a prediction, recommendation or classification (Section 2, term 4), including generative AI (Section 3).
A Provider develops an AI system, or has one developed, and puts it into service under its own name (Section 2, term 17). A User operates a system under its authority (Section 2, term 27) and becomes a Provider if it materially changes a model and offers it under its own name (Section 14.3). A substantially modified system counts as a new one (Section 2, term 13).
What kind of rule it is
It is a supervisory guideline from QCB. Its text calls its contents instructions (Section 1). Most duties are written as must, but some are written as should or may, and this page says so where it matters. It has no penalty clause of its own, but Entities must comply with it together with the QCB Law (Section 24.1), which lets QCB fine institutions for breaching instructions issued under that law (QCB Law, Article 216). An Entity that cannot meet a requirement asks QCB for an exemption, backed by a business case and recorded with an expiry date, and stays accountable for the risk (Sections 23.1 to 23.6).
High risk AI and the three oversight protocols
High risk AI is a system that could significantly harm the Entity's operations or the financial system, or limit people's access to financial resources or essential services (Section 2, term 10). Whatever the Entity's own assessment, a system is high risk when it determines consumer access to financial services, materially affects employees, or processes sensitive personal information such as health or children's data (Section 9.7). Systems without direct human oversight will probably rate higher (Section 9.5.1).
Every AI system needs one of three oversight protocols (Section 13.1). The guideline does not say human in the loop; the table maps its terms to those of the National Cyber Security Agency, NCSA (NCSA Guidelines, Section 5.1.5). For high risk AI the Supervisor can override outputs and stop the system (Section 13.4).
| QCB protocol | Meaning | NCSA term | Main duties |
|---|---|---|---|
| AI assisted decision making | A human approves designated outputs | Human in the loop | Trained operators (13.8.1) |
| Human exception oversight | A human monitors and can take over or shut down | Human over the loop | Takeover tools and limits the system cannot override (13.7) |
| Fully autonomous AI | The algorithm runs normal operations | Human out of the loop | Detailed case for QCB approval (13.5.1); if high risk, prior approval, early notice, guard rails, limits and a shutdown option (13.6) |
QCB approval, reporting and the AI register
QCB may send a system to a sandbox before approving it (Section 11.3). The register covers every AI arrangement (Section 10.1) and records per system the high risk rating, the Entity's role as User or Provider, a functional category and the oversight protocol (Section 10.7). For bought or outsourced systems it adds the Provider, a supplier assessment, governing law and contract dates, and for high risk systems how replaceable the system is (Section 10.7.5).
| Trigger | What QCB needs | Section |
|---|---|---|
| New AI system launched as Provider, or a material change | Approval before launch | 11.1 |
| Buying, licensing or outsourcing high risk AI | Approval before signing | 11.2 |
| Outsourcing any AI activity | Prior consent and regular due diligence (written as should) | 12.1 |
| New high risk AI system | Training, validation and test results for approval | 14.1 |
| Using a Provider's system | Provider instructions and results, own tests and oversight plan | 14.2 |
| Fully autonomous high risk AI | Prior approval and notice once it is being considered | 13.6.1 |
| Serious incident | Report immediately once the Provider has established a causal link or its reasonable likelihood (written as should) | 19.5 |
| Register | Every year and on request | 10.6 |
Governance, data, security and records
The board and senior management remain accountable for AI outcomes, including decisions made on the Entity's behalf (Section 7.1). The board sets the tolerated AI exposure and assigns accountability (Section 7.2), and senior management needs at least one member who understands technology risk (Section 7.3.1). The Entity keeps an AI strategy (Section 6.1) and committees that assess use cases before implementation (Section 7.5). A User or Provider of high risk AI puts a written framework in place covering compliance, design, testing, data management, monitoring, record keeping and accountability (Section 16.1).
Training, validation and test data should be kept separate (Section 15.3), accuracy must be measured on test data and models must be tested for bias across demographic groups (Sections 15.6 and 15.7). Providers hand over development and testing data by contract, and the Entity makes it available to QCB (Sections 15.1 and 15.2). An AI trust, risk and security management (AI TRiSM) program covers content anomaly detection, data protection from third parties, an acceptable use policy and bias assessments (Section 17.2); models are protected against manipulation, query attacks and prompt injection (Sections 17.5 to 17.7); and every Entity using AI deploys a data loss prevention tool and content anomaly detection tools (Sections 17.8 and 17.9). Records should include audit logs with traceability of decisions, a record of each model version and archived data sets (Section 19.3).
Duties toward customers
An Entity must tell customers when they interact with an AI system (Section 20.1). It should also inform them of the AI, its risks and limits: for systems that are not high risk, such as chatbots, before first use, for example in the terms of use (Sections 20.3.1 and 20.3.3); each update counts as a new version needing notice (Section 20.3.2); for high risk AI such as credit or insurance, at every interaction (Section 20.3.4). Customers must get instructions and, on request, an explanation of the data and factors behind a decision, and be told how a decision may affect them and whether it can be reversed (Sections 20.4 to 20.6). High risk systems prompt customers to check that their data is up to date, and every Entity provides a feedback channel (Sections 20.7 and 20.8). The Entity should collect consent to the risks of AI before the service starts (Section 20.9). It need not disclose intellectual property or source code, and AI used to detect fraud and identity theft is outside these disclosure duties (Sections 20.10 and 20.11).
Customers can query AI decisions and ask for review of decisions taken without a human (Section 21.1). After a negative decision they choose between resubmitting with new data and review by a qualified human (Section 21.2). An opt out is not mandatory: the Entity weighs harm, reversibility and feasibility, and if it offers none, must provide another route of recourse (Sections 22.1 and 22.2). Later complaints go through the standard complaint process (Section 21.3).
Outsourcing, third party AI and data location
Before outsourcing, the Entity must run due diligence on the provider's capabilities (Section 12.2), should assess where data is processed, stored and transmitted and which subcontractors are involved (Section 12.3), and must get board approval (Section 12.5). Its auditors must be able to review the provider's accounting records and internal controls (Section 12.8), it keeps a contingency plan for a sudden end of the arrangement (Section 12.9), and it stays responsible for the provider's acts (Section 12.11). A service agreement is mandatory, and it should contain at least the clauses listed below (Section 12.10).
The location rules sit in companion regulations. Personally identifiable and financial information must be processed within Qatar only, and cloud arrangements need prior QCB approval (Cloud Computing Regulation, Sections 21.4 and 21.5); that regulation allows no departure except through a QCB exemption (Section 23.1). The primary environment for personal and sensitive financial data must be in Qatar, unless QCB gives explicit approval, and such data may not be stored or transferred abroad without QCB approval (Data Handling and Protection Regulation, Sections 7.6, 7.7 and 15.1). For an AI agent this reaches the platform, the model endpoint, speech services and logs.
- Right to monitor the provider's compliance and require remediation (12.10.1)
- Ethics, defined roles, confidentiality, security and business continuity (12.10.2 to 12.10.5)
- Right to terminate on breach or QCB direction (12.10.6)
- Data wiped on termination (12.10.7)
- QCB right to audit the provider's accounts (12.10.8)
Who has to do what
Entity (QCB regulated financial institution)
The bank or insurer that puts the AI agent in front of customers or staff.
- Keep an AI strategy and classify each system. (Sections 6.1 and 9.7)
- Keep the AI register and send it to QCB yearly. (Sections 10.1 and 10.6)
- Get board approval before outsourcing AI and, as the guideline says it should, QCB's prior consent. (Sections 12.1 and 12.5)
- Tell customers they deal with AI and offer human review. (Sections 20.1 and 21.2)
Board of Directors and senior management
The board and executives who answer for AI use.
- Remain accountable for AI outcomes and decisions. (Section 7.1)
- Approve the tolerated AI exposure and assign accountability. (Sections 7.2.1 and 7.2.3)
Provider
Whoever puts an AI system into service under its own name, which includes a bank launching a branded agent.
- Get QCB approval before launch or a material change. (Section 11.1)
- Manage the system over its life cycle and share its data by contract. (Section 15.1)
User
The Entity running a system it bought or licensed, such as a language model.
- Submit the Provider's instructions and results plus its own tests and oversight plan. (Section 14.2)
- Assign a trained Supervisor with authority. (Section 13.3)
Outsourcing service provider
The platform vendor, model provider and speech providers, bound through the Entity's contract.
- Accept the minimum contract clauses, QCB audit and data deletion at exit. (Section 12.10)
- Follow the Entity's AI TRiSM program. (Section 17.3)
Supervisor and Operator
Staff who monitor the agent, approve actions or take over.
- Be given the tools and authority to override, reverse or stop a high risk system. (Sections 13.4.3 and 13.4.4)
- Be trained to interpret AI output used for decisions. (Section 13.8.1)
What it means for chatbots, voice agents and agents that act
A customer facing chatbot is the guideline's example of a system that is not high risk (Section 20.3.1). It still goes on the register (Section 10.1), needs an oversight protocol (Section 13.1), an AI notice (Section 20.1) and, when launched under the bank's name, QCB approval (Section 11.1). Once it processes sensitive personal information or determines access to a product, it is high risk (Section 9.7).
A voice agent carries the same duties on the call itself (Section 20.1). Recordings and transcripts hold personal and financial data, so cloud speech and model services must process them in Qatar (Cloud Computing Regulation, Section 21.4), and storing or sending them abroad needs QCB approval (Data Handling and Protection Regulation, Section 15.1).
An agent that makes payments or changes accounts without human oversight is fully autonomous AI: it needs a detailed case for QCB approval (Section 13.5.1), QCB approval before data is processed that way (Data Handling and Protection Regulation, Section 13.5) and, if high risk, guard rails and a shutdown option (Section 13.6). Human approval per action makes it AI assisted, which rates lower (Section 9.5.1).
An internal staff assistant is high risk when it supports decisions that materially affect employees (Section 9.7.2). Staff need training on AI limits (Section 8.2.5) and the data loss prevention duty applies (Section 17.8).
How the Blits.ai platform helps
The organization that puts an AI agent in front of its customers or staff stays responsible for meeting the rule. These are the parts the platform takes care of or records for you.
Tell customers they deal with AI (Section 20.1)
The AI notice shows in the chat that content is AI generated, set per language in system messages, and a voice agent can say so at the start of a call.
Human oversight, override and stop (Sections 13.1, 13.4 and 13.7)
Each tool is automatic, requires approval or disabled; tools that move money or change data require approval by default, and approvers see the exact frozen tool call. A global stop halts outbound campaigns.
See how it worksBuilt in limits (Sections 13.6.2 and 13.7.5)
Guardrails with rules, blocked topics and a judge model check questions and answers per agent; approvals can apply above an amount, and spend caps limit usage.
See how it worksAudit logs, traceability and versions (Sections 19.3.1 and 19.3.3)
A decision log per conversation records tool calls, guardrail checks, approvals and the model used, exportable as JSON or streamed to a SIEM. Version history with rollback covers assistants and prompts.
See how it worksTesting and monitoring (Sections 15.6 and 19.1)
Test suites and scheduled monitors run conversations against an agent and report regressions, and a dry run tests agentic tasks before they act.
AI security and personal data (Sections 17.2 and 17.9)
A content scanner runs on every assistant, guard models run inside the platform, and the Privacy Shield masks personal data before it is stored or sent to a model.
See how it worksHuman review and consent (Sections 20.9 and 21.2)
Conversations can be handed to a live agent in Salesforce, Zoho SalesIQ or Freshdesk. Consent can be asked before a conversation starts, is recorded and exports as CSV.
See how it worksData processed within Qatar (Cloud Computing Regulation Section 21.4; Data Handling and Protection Regulation Section 15.1)
Blits.ai runs a managed region in Qatar and deploys in private cloud or on premises. A tenant can limit which AI providers receive its data and connect its own model endpoints.
See how it worksVendor due diligence (Sections 9.5.2 and 12.2)
Blits.ai holds ISO 27001, SOC 2 Type II, PCI DSS Level 1 and ISO/IEC 42001 certifications. QCB approvals, the register and contract terms stay with the Entity that deploys the agent.
See how it works
Fines and other consequences
- Up to QAR 10,000,000 per violation of the law or instructions issued under it, and up to QAR 100,000 a day for a continuing violation, after a warning. (QCB Law No. (13) of 2012, Article 216)
- Up to QAR 2,000,000 for refusing QCB data or access, or giving it misleading information. (QCB Law No. (13) of 2012, Article 217)
- Under the PDPPL, up to QAR 1,000,000 for breaching consent, notice or controller duties, and up to QAR 5,000,000 for breaching security or special nature data rules. (PDPPL, Articles 23 and 24)
- In the QFC, up to USD 1,500,000 per infringement of the Data Protection Regulations. (QFC Data Protection Regulations 2021, Article 36(3))
Questions about QCB Artificial Intelligence Guideline
Sources
- Qatar Central Bank: Artificial Intelligence Guideline (PDF)
- Qatar Central Bank: Cloud Computing Regulation (PDF)
- Qatar Central Bank: Data Handling and Protection Regulation (PDF)
- Qatar Central Bank: Technology Risks, January 2018 (PDF)
- Qatar Central Bank: Instructions to Banks, Part XI, Sanctions and Financial Penalties, with QCB Law Chapter 9 (PDF)
- NCSA: Law No. (13) of 2016 on Personal Data Privacy Protection, unofficial English text (PDF)
- NCSA: Data Privacy Impact Assessment, Guideline for Regulated Entities (archived copy)
- NCSA: Personal Data Breach Notifications, Guideline for Regulated Entities (archived copy)
- NCSA: Guidelines for Secure Adoption and Usage of Artificial Intelligence, version 1.0 (archived copy)
- QFC Data Protection Office: Data Protection Regulations and Rules 2021 Guidance, version 3.1 (PDF)
- Qatar Center for Artificial Intelligence: National Artificial Intelligence Strategy for Qatar (PDF)
- Qatar Center for Artificial Intelligence: adoption of the National AI Strategy
Read next
AI use cases QCB Artificial Intelligence Guideline applies to
No use case in the library refers to it yet. Browse the AI Use Case Library.