Frequently asked questions about the National AI Risk Management Framework

This FAQ gives short information about the National AI Risk Management Framework (SDAIA-P145) of SDAIA and about its use in Modulos. Each part of this FAQ refers to the section with the full information.

Disclaimer

This page gives general information. It is not legal advice.

What is the SDAIA National AI Risk Management Framework?

SDAIA-P145 is SDAIA's advisory framework for AI risk management in government and private-sector entities (p. 6). The full framework is in Arabic, version 1, April 2026. It has four reference pillars and five steps: context and scope, risk identification, risk assessment, risk treatment, and monitoring and review (Figure 3, p. 13).

See the framework and its structure.

Is the SDAIA AI Risk Management Framework mandatory?

SDAIA gives SDAIA-P145 (its National AI Risk Management Framework) as an advisory framework (p. 6). SDAIA-P145 helps entities write internal policies and procedures, and it does not set detailed technical standards for model performance (p. 8).

In its reference pillars, SDAIA-P145 tells entities that AI regulations set obligations and responsibilities for the related entities. It also tells entities that data regulations put an obligation on entities to apply lawful and safe practices (p. 16). SDAIA-P145 also tells entities that the regulatory requirements, policies, and guidelines of each sector control the design, development, and operation of AI systems in different contexts of use. AI risks change with the sector, the sensitivity of the data, the possible impact scope, and the type of decisions, which makes special controls and procedures necessary (p. 16).

See the framework and the regulations.

Who is the SDAIA AI Risk Management Framework for?

SDAIA-P145 (SDAIA's National AI Risk Management Framework) is for all government and private-sector entities in all sectors and at all levels of digital maturity (p. 8). Entities can use it with flexibility for different usage patterns and magnitude of risks. The condition is that they keep the principles of systematic risk management. In its Scope section, SDAIA-P145 focuses on AI systems and applications with high impact on persons, assets, and operations (p. 8).

SDAIA-P145 names three groups: system developers, system operators, and policy makers (p. 7). Policy makers examine if the available regulatory frameworks are prepared for AI risks. They identify where AI risk governance is not sufficient, and make policies and legislation with different levels. SDAIA-P145 does not give this work to the persons who write internal policies.

See Entities and the three groups.

Is the SDAIA framework available in English?

Only an executive summary is in English: SDAIA-P145EN, version 1.0, April 2026. The full framework is only in Arabic: SDAIA-P145, version 1, April 2026. The English executive summary is not the full framework. If the two are different, use the Arabic framework.

The SDAIA Publications entry gives the two PDFs with the date "1/4/2026". This is not a date from which the framework applies.

See the framework sources.

What are the four reference pillars?

The four reference pillars of SDAIA-P145 (SDAIA's National AI Risk Management Framework) are general principles, AI regulations, data regulations, and sector regulations (p. 13–14). They are a governing reference for the design, development, and operation of AI systems (p. 14). SDAIA-P145 tells entities that it is necessary to think about the reference pillars when they make and use AI systems (p. 14). General principles include Islamic values and seven AI ethics principles.

See The four reference pillars.

What are the seven AI ethics principles?

SDAIA-P145 (SDAIA's National AI Risk Management Framework) gives these seven AI ethics principles in the general principles reference pillar (p. 14–15):

1. Integrity and fairness
2. Privacy and security
3. Humanity
4. Social and environmental benefits
5. Reliability and safety
6. Transparency and explainability
7. Accountability and responsibility

SDAIA-P145 tells entities that it is necessary to think about these AI ethics principles through all of the system lifecycle (p. 14).

See General principles and AI ethics.

What does the context-and-scope stage cover?

In SDAIA-P145 (SDAIA's National AI Risk Management Framework) the context and scope step is the first of the five steps (p. 17). It records these items:

• Direct users and indirectly affected groups.
• Data flows, input points, output points, and all value-chain parties with their roles and responsibilities.
• The automation level and the human role in decisions: the points of approval, human review, objection, and stop, and when human intervention is mandatory.
• The responsibilities of developers, operators, and decision owners, the formal approval paths, and the escalation mechanisms.
• The lifecycle and change management.

SDAIA-P145 also recommends that entities think about a clear identification of intended uses against unauthorized uses and out-of-scope uses (p. 17).

For changes, SDAIA-P145 recommends that entities think about two items (p. 17). The first item is the requirements for a new risk assessment after material changes. The other item is the mechanisms that tell users or affected groups about changes that can have an effect on performance or risk.

See Context and scope.

What should an internal AI risk policy cover?

SDAIA-P145 (SDAIA's National AI Risk Management Framework) helps entities write internal policies and procedures for risk identification, risk assessment, and risk treatment or acceptance, which fit the conditions of each entity (p. 8). The Scope section includes governance, risk scenario records, an acceptable risk level and its controls, and mechanisms for monitoring and response. SDAIA-P145 adapts to different usage patterns and to the magnitude of risks, on the condition that entities keep the principles of systematic risk management (p. 8).

See Internal AI risk policies and procedures and Entities.

When can an entity accept an AI risk?

In SDAIA-P145 (SDAIA's National AI Risk Management Framework) acceptance is one of the four treatment options (p. 26). Acceptance applies when the residual risk is shown to be in the approved risk tolerance, after a check shows that the minimum controls are available. The acceptance record gives the reasons for acceptance, the limits of use, the conditions, and the review plan (p. 26).

The introduction to risk treatment gives an example, not a new condition: acceptance of a risk of small impact that agrees with the approved risk tolerance (p. 26).

Other provisions of SDAIA-P145 address the decision:

• The acceptance or escalation decision uses the residual-risk reassessment results, the approved risk tolerance, and the context of use. High risks go to the applicable governance level (p. 27).
• SDAIA-P145 recommends that entities think about a decision level that agrees with the risk level of the system (p. 27).

See Treatment strategy selection and Acceptance or escalation decision.

What risk records does the SDAIA framework describe?

SDAIA-P145 (SDAIA's National AI Risk Management Framework) includes the task to record risk scenarios (p. 8). It also helps entities use the same names and methods to record and compare risks (p. 18). It sets no one location for all records. The record provisions give these relations (p. 26–29):

• Risk treatment decisions and periodic review outputs: in the risk register.
• Residual-risk reassessment results: recorded and connected to the risk register and treatment plans.
• Continuous performance monitoring results: connected to the risk register.
• Incidents, root-cause analysis results, and lessons learned: recorded and connected to the risk register.

The example records apply to one example (p. 34–36).

Does SDAIA require risk-specific AI training?

SDAIA-P145 (SDAIA's National AI Risk Management Framework) does not set one training plan or one training interval for all entities. It helps entities to be better prepared, to increase human capacity, and to increase risk awareness (p. 7). Its example includes user training on model limits, data classification, and privacy (p. 42), and measures to keep human competence (p. 44).

These measures apply to that example only. The risk inventory includes the possible decrease of cognitive skills with time (p. 19). Policy makers have a different function for regulatory frameworks (p. 7).

See Prepare personnel and The risk inventory.

What are SDAIA's seven AI risk categories?

SDAIA-P145 (SDAIA's National AI Risk Management Framework) recommends «يُقترح» a unified risk taxonomy for its risk inventory and tells entities that they can use these seven risk categories (p. 18–20):

1. Bias, discrimination, and abuse
2. Privacy and security
3. Misinformation
4. Malicious use
5. Human-machine interaction
6. Social, economic, and environmental impacts
7. Safety and limitations

Entities can use this risk taxonomy for a complete risk inventory of technical risks and non-technical risks. The records are clear, and entities can compare them.

See The risk inventory.

Can we use our own AI risk taxonomy?

SDAIA-P145 (SDAIA's National AI Risk Management Framework) gives no procedure to accept a different risk taxonomy in its risk inventory task (p. 18). It recommends «يُقترح» a unified risk taxonomy and tells entities that they can use its seven risk categories. The risk inventory task helps entities prepare a complete risk inventory with records that they can compare.

See The risk inventory.

How is risk likelihood estimated?

In SDAIA-P145 (SDAIA's National AI Risk Management Framework) the estimate of likelihood is for the approved period (p. 22). It uses the available indications, the related data, and practical experience of how the system and its operating environment operate (p. 22). The four levels are rare (1), unlikely (2), likely (3), and almost certain (4).

SDAIA-P145 recommends that entities think about one or more likelihood factors that apply to the risk and the system context. The likelihood factors are exposure, preventive controls, ease of occurrence, and precedents. It is not necessary to use all of them each time. Strong preventive controls decrease likelihood. More exposure, more ease of occurrence, and more precedents increase it (p. 22–23).

The estimate does not initially use specified percentages. The levels agree with the annual reference ranges that some risk management frameworks use (p. 23).

When is catastrophic impact used?

SDAIA-P145 (SDAIA's National AI Risk Management Framework) keeps catastrophic impact only for extreme cases: very severe, very dangerous harm with wide and deep effects (p. 23). The harm can be irreversible, or a long recovery can be necessary, with permanent or almost permanent effects.

The typical properties of catastrophic impact include these properties (p. 23–24):

• An effect directly on public security or societal stability, or threats of national dimensions.
• An effect directly on human safety, or severe risk to life, physical health, or psychological health.
• Harm to religion, offense to sacred things, religious incitement, or religious misinformation, with possible severe and sensitive social effects.
• Actual deprivation of fundamental rights or vital services that cannot be compensated for easily or in a short period.
• Breach or control of critical systems, or unauthorized access to very sensitive basic components for operation.
• Wide leakage of very sensitive data.

Catastrophic impact usually occurs when the risk has these properties. They are not the only conditions. SDAIA-P145 does not tell entities if one property is sufficient or if all the properties are necessary.

How does SDAIA's 4×4 risk matrix work?

In SDAIA-P145 (SDAIA's National AI Risk Management Framework) the risk level is the likelihood of the risk multiplied by its impact. Likelihood and impact each have levels 1–4, which give risk levels from 1–16 (p. 24–25). Likelihood uses the approved period (p. 22). The four risk groups are low (1–2), medium (3–6), high (8–12), and catastrophic (16).

Only almost certain likelihood with catastrophic impact gives 16. Catastrophic impact is for extreme cases (p. 23). System classification is different. These provisions give no method to calculate it from the risk levels of the risks (p. 7, 24–25, 28).

See Risk level and risk matrix.

What are the four risk treatment strategies?

SDAIA-P145 (SDAIA's National AI Risk Management Framework) gives four treatment options: avoidance, mitigation, risk transfer, and acceptance (p. 26). Before the selection, the entity examines each risk and its risk assessment, and sets the treatment objective. The entity records each decision at the level of the risk register. The record gives the reasons, the responsibilities, the approval and escalation requirements for high risks, and measurement and follow-up indicators. An audit can examine each decision.

• Avoidance: for a high risk that it is not possible to decrease to an acceptable risk level.
• Mitigation: technical controls, organizational controls, and operational controls. Control of permissions is one example.
• Risk transfer: for example contracts, insurance, or obligations on suppliers. The entity keeps its governance and compliance responsibility.
• Acceptance: residual risk shown to be in the approved risk tolerance, after a check of the minimum controls. A record gives the reasons, use limits, conditions, and a review plan.

For the full treatment options, see Treatment strategy selection (p. 26). For the controls, see Risk treatment controls (p. 26).

How is residual risk reassessed?

In SDAIA-P145 (SDAIA's National AI Risk Management Framework) after controls are applied, the entity estimates likelihood and impact again with the same approved risk assessment method (p. 27). It uses the applied controls and the changes to the system design or to the operating environment. The result shows if the risk decreased as necessary. It also shows if residual risk is in the approved risk tolerance or makes more measures or escalation necessary.

Results are recorded and connected to the risk register and treatment plans (p. 27). See Residual-risk reassessment.

Who makes the final operating decision?

SDAIA-P145 (SDAIA's National AI Risk Management Framework) names no one person to make the decision. It recommends that entities think about two items (p. 27). These items are a decision level that agrees with the risk level of the system, and clear records of approvals and escalation paths.

The decision uses the residual-risk reassessment results and the approved risk tolerance. It also uses how sensitive the use case is, the scope of impact, and the type of decisions that the system helps to make or makes. The decision is acceptance when residual risk is in the acceptable risk limits. High risks go to the applicable governance level for conditional release, restricted use, or a temporary or permanent stop (p. 27).

Context and scope identifies the responsibilities of developers, operators, and decision owners, the formal approval paths, and the escalation mechanisms (p. 17).

What does monitoring and review cover?

In SDAIA-P145 (SDAIA's National AI Risk Management Framework) monitoring and review includes performance checks, periodic reviews, and incident response (p. 28–29):

• Performance checks: data drift and model drift, output quality, beneficiary experience, and unexpected system behavior that can have an effect on fairness, safety, or compliance. Fairness indicators and bias indicators apply where available. Alert thresholds have response plans, and results connect to the risk register (p. 28).
• Periodic reviews: scheduled or event-based reviews, with no specified interval. Data or model changes, an increase in the scope of use, and new regulatory requirements are examples. Outputs go in the risk register (p. 28).
• Incident response: report incidents or risk indicators immediately when they are found, with escalation by severity and scope of impact (p. 29).

When should an AI incident be reported?

SDAIA-P145 (SDAIA's National AI Risk Management Framework) includes mechanisms to report incidents or risk indicators immediately when they are found (p. 29). It gives no time limit as a number and names no external organization that gets the reports.

Incident response includes escalation by severity and scope of impact to include competent parties and decision makers in time. Root-cause analysis identifies the causes of incidents and examines if controls are sufficient, with corrective actions and preventive actions. Incident records, root-cause analysis results, and lessons learned connect to the risk register. These records adjust likelihood and impact estimates and increase control and policy quality (p. 29).

How do OFF-23 and MFF-23 divide the work in Modulos?

In Modulos, Saudi AI Risk Management has an Organization-level Framework, OFF-23, and an AI System-level Framework, MFF-23. Modulos recommends OFF-23 on the organization AI System. It recommends MFF-23 on each AI System or use case to which the entity applies SDAIA-P145. Modulos recommends this structure. If your organization has no organization AI System, contact Modulos before you add OFF-23.

See Recommended AI System structure, Requirements, and Modulos limits.

In what order should we set up Saudi AI Risk Management in Modulos?

Modulos recommends that you prepare governance with OFF-23 first. If approved risk tolerance and risk-record procedures are available, refer to them. For each MFF-23 AI System, start with reference pillars, context and scope, risk identification, and risk assessment. Then examine the risk, set the treatment objective, select treatment options and controls, estimate residual risk again, and make the decision. Monitoring and review continue during operation. See Work sequence.

The Requirement Owner examines completed Controls before the Requirement Owner marks the Requirement as Fulfilled. See Requirement decisions and Modulos limits.

Does Modulos record risk treatment and acceptance decisions?

You cannot record risk treatment decisions with the risk functions of Modulos, and they do not record the acceptance of residual risk. Modulos recommends that you record these decisions in your work on the Requirements MRF-424 to MRF-427. The Frameworks help you record and examine work on SDAIA-P145 (SDAIA's National AI Risk Management Framework). See Modulos limits for the work that these risk functions do not do.