NIST AI RMF DATA INFRASTRUCTURE MODEL

Categorization and risk tiering

Part of: MP: MAP — Context & Risk Identification

Description

AI systems are categorized based on their intended use, beneficiaries, and potential for harm, enabling risk-proportionate controls. Not all AI systems carry equal risk — treating them equally wastes resources on low-risk systems and under-protects high-risk ones.

Suggested Actions

1

Develop an AI system classification framework based on: autonomy level (advisory vs. autonomous), decision impact (reversible vs. irreversible), affected population (internal vs. public), and domain sensitivity (entertainment vs. healthcare)

2

Assign risk tiers using a consistent methodology: Minimal (standard controls), Limited (enhanced monitoring and transparency), High (full TEVV, human oversight, fairness audits), Unacceptable (prohibited)

3

For each risk tier, define the minimum required controls: testing requirements, human oversight level, monitoring frequency, documentation standards, and approval authority

4

Categorize systems by their potential for discriminatory impact — AI systems that influence decisions about individuals (hiring, lending, insurance, benefits) require the highest scrutiny regardless of other risk factors

5

Re-assess risk tiers when system scope expands, user population changes, or the system is applied to new domains — scope creep is a common vector for risk escalation

6

Document classification rationale for every system with enough detail that an auditor can understand and reproduce the assessment

Related Controls

AI Risk Self-Assessment AI Red Teaming and Adversarial Testing