Introduction to NIST AI RMF

The foundations of the AI Risk Management Framework and how to begin implementation.

1. What is NIST AI RMF?

The NIST AI Risk Management Framework is a voluntary framework developed by the US National Institute of Standards and Technology to manage risks associated with AI systems.

Key characteristics

Characteristic	Description
Voluntary	Not legally binding; a recognised source of good practice
Flexible	Adaptable to different organisation sizes and sectors
Iterative	Supports a continuous improvement cycle
Complementary	Works alongside existing standards and frameworks, including ISO standards and NIST CSF

Versions and publications

Publication	Date	Purpose
AI RMF 1.0	January 2023	Core framework
AI RMF Playbook	January 2023: partial release; 30 March 2023: first complete version	Suggested practical actions
AI 600-1, Generative AI Profile	July 2024	Considerations specific to generative AI

2. Why use NIST AI RMF?

For organisations in Czechia

Reason	Explanation
EU AI Act mapping	NIST practices can be mapped to relevant EU AI Act requirements
International recognition	Relevant to organisations with US parent companies or customers
Practical guidance	Voluntary actions selected for the organisation’s context
Established practice	A recognised reference in the AI governance community

Working alongside EU rules

Methodology and law have different roles

NIST AI RMF helps structure risk management. The AI Act establishes obligations according to the organisation’s role, the system and the applicable dates. Selected measures may serve more than one purpose, but a mapping alone does not demonstrate compliance or programme completeness. The NIST Playbook is a voluntary collection of suggestions, not a mandatory checklist to implement in full.

3. The four core functions

3.1 GOVERN — Governance

Purpose: Establish the culture and processes needed to manage AI risk.

Key areas:

• Policies and procedures
• Roles and responsibilities
• Organisational culture
• Third-party management

Possible outputs:

• AI governance policy
• Defined roles, such as an AI Risk Officer or AI Ethics Board where appropriate
• Risk tolerance thresholds
• Vendor management processes

3.2 MAP — Mapping

Purpose: Identify the context, intended purpose and risks of AI systems.

Key areas:

• AI system inventory
• Definition of intended use
• Stakeholder identification
• Impact assessment

Possible outputs:

• AI system inventory
• Risk classification
• Stakeholder map
• DPIA or AI impact assessment documentation, where relevant

3.3 MEASURE — Measurement

Purpose: Analyse, test and validate AI systems.

Key areas:

• Testing, evaluation, verification and validation (TEVV)
• Metrics and KPIs
• Bias assessment
• Security testing

Possible outputs:

• Test results
• Performance metrics
• Fairness reports
• Security assessment

3.4 MANAGE — Management

Purpose: Prioritise risks and implement mitigation measures.

Key areas:

• Risk prioritisation
• Mitigation actions
• Incident management
• Continuous improvement

Possible outputs:

• Risk treatment plans
• Incident response procedures
• Remediation actions
• Improvement roadmap

4. Characteristics of trustworthy AI

NIST describes seven characteristics of trustworthy AI.

4.1 Accountable and transparent

Questions to answer

• Who is accountable for the AI system?
• Are decisions documented?
• Is the process transparent to stakeholders?
• Is there an audit trail?

Practical steps:

1. Assign an AI System Owner to each system.
2. Implement logging and an audit trail.
3. Document decision-making processes.
4. Communicate clearly about AI use.

4.2 Fair, with harmful bias managed

Questions to answer

• Are outcomes fair across demographic groups?
• Is training data tested for bias?
• Are disparate impacts monitored?
• Is there human oversight of sensitive decisions?

Practical steps:

1. Test for bias before deployment.
2. Monitor fairness metrics in production.
3. Establish an appeal or review process.
4. Involve diverse perspectives in evaluation teams.

4.3 Privacy-enhanced

Questions to answer

• Is personal data minimised?
• Is there an appropriate legal basis, including consent where applicable?
• Is data anonymised where feasible?
• Are retention periods defined?

Practical steps:

1. Assess whether a DPIA is required under GDPR Article 35 and the risks of the processing. Where high risk is likely, conduct it before processing begins.
2. Implement privacy by design.
3. Minimise data collection.
4. Address the applicable GDPR obligations.

4.4 Safe

Questions to answer

• Could the system cause physical harm?
• Could it cause psychological harm?
• Are there fail-safe mechanisms?
• Have edge cases been tested?

Practical steps:

1. Identify potential harm scenarios.
2. Implement appropriate safety safeguards.
3. Test adversarial inputs.
4. Provide a human override capability.

4.5 Secure and resilient

Questions to answer

• Is the system protected against cyberattacks?
• Is it resilient to adversarial attacks?
• Is there an incident response plan?
• Is business continuity addressed?

Practical steps:

1. Conduct a security assessment.
2. Test adversarial robustness.
3. Implement monitoring and alerting.
4. Prepare incident response procedures.

4.6 Valid and reliable

Questions to answer

• Does the system work correctly in its intended context?
• Are its outputs consistent?
• Are accuracy and performance measured?
• Is the system revalidated regularly?

Practical steps:

1. Define performance metrics.
2. Implement continuous monitoring.
3. Revalidate the system regularly.
4. Document limitations.

4.7 Explainable and interpretable

Questions to answer

• Can you explain how the system works?
• Can you explain a specific decision?
• Is the explanation understandable to end users?
• Are the factors affecting outputs known?

Practical steps:

1. Document the model architecture.
2. Implement suitable explainability tools.
3. Prepare explanations that users can understand.
4. Assess the information duties and rights associated with automated decision-making under GDPR Articles 13–15 and 22. Their scope depends on the processing involved.

5. Getting started

The weeks below illustrate an internal implementation plan. They are not a promised service delivery time.

Step 1: Create an AI inventory — week 1

Record the following for each AI system:
- [ ] Name and ID
- [ ] Intended purpose
- [ ] Owner
- [ ] Type: internal / third-party / embedded
- [ ] Status: development / production / retired
- [ ] Risk level and the basis for classification

Keep internal risk grades separate from legal classifications under the AI Act.

Step 2: Define responsibilities — week 2

Role	Responsibility	Who fills this role in your organisation?
AI System Owner	Overall responsibility for the system
AI Risk Contact	Risk management and compliance coordination
Technical Lead	Technical implementation and testing
Data Steward	Data quality and privacy

Step 3: Classify risks — week 3

First identify the system’s purpose and your role. Then assess the prohibitions in Article 5 and high-risk classification under Article 6(1) and Annex I, and Article 6(2) and Annex III, including relevant exceptions. Assess Article 50 transparency duties, any obligations concerning GPAI models and other applicable rules separately. Record the evidence, responsible person and applicable deadline for each conclusion. See the classification guide.

Step 4: Implement basic controls — week 4

Suggested starting controls to adapt:

• Documented purpose and use
• Named owner
• Basic monitoring
• User awareness

For high-risk systems, check the obligations attached to your role. A provider should address, in particular:

• Risk management system
• Data governance
• Technical documentation
• Human oversight
• Accuracy and robustness testing

6. Indicative self-assessment

The following eight-question assessment is a TECHNOMATON discussion aid. It is not a validated diagnostic tool, an official NIST score or a legal compliance assessment. Consider the evidence and differences between individual systems. A low result in a critical area cannot be offset by a high average elsewhere.

Self-assessment questions

Rate your organisation from 1 to 4:

Area	Question	Score
GOVERN	Have we defined AI policies and roles?	/4
GOVERN	Is AI governance part of our organisational culture?	/4
MAP	Do we have an inventory of all AI systems?	/4
MAP	Do we conduct impact assessments?	/4
MEASURE	Do we test AI before deployment?	/4
MEASURE	Do we monitor performance in production?	/4
MANAGE	Do we have incident response procedures for AI?	/4
MANAGE	Do we improve our approach continuously?	/4

Interpretation

Average score	Indicative position	Next step
1 to 1.5 inclusive	Getting started	Clarify purpose, inventory and responsibilities
Above 1.5 to 2.5 inclusive	Establishing a foundation	Fill gaps in rules and measurement
Above 2.5 to 3.5 inclusive	Repeating the process	Verify evidence, exceptions and corrective actions
Above 3.5 to 4	Improving continuously	Check control effectiveness and changing risks

Partial, Risk Informed, Repeatable and Adaptive are Tiers in the NIST Cybersecurity Framework. They are not NIST AI RMF maturity levels or mathematical score bands.

7. Common mistakes

Mistake	Consequence	How to avoid it
No inventory	Unknown risks	Start with MAP
Unclear responsibilities	Nobody is accountable	Define roles through GOVERN
Missing testing	Unreliable systems	Implement TEVV
Reactive approach	Late incident response	Use MANAGE proactively
Isolated activities	Inconsistent governance	Integrate with existing processes

8. Next steps

Continue with:

1. GOVERN — Implementing governance
2. MAP — Creating an AI inventory and risk assessment
3. MEASURE — Testing and measuring AI systems
4. MANAGE — Risk and incident management
5. Generative AI risks — Generative AI considerations

References

Specialist references checked on 6 September 2026.

• NIST: Cybersecurity Framework Tiers
• NIST AI RMF 1.0
• NIST AI RMF Playbook
• NIST AI 600-1 Generative AI Profile

AI-Native Entry Framework | CC BY-NC-SA 4.0