AI Governance

AI risk management with NIST AI RMF and ISO/IEC 42001, the EU AI Act risk tiers, AI use-case registries, model documentation, human oversight and responsible AI operating models.

Frameworks and regulation

Item Summary
NIST AI Risk Management Framework (AI RMF 1.0, 2023) Functions: Govern, Map, Measure, Manage; plus a Generative AI profile (2024)
ISO/IEC 42001:2023 AI management system standard (certifiable), similar in spirit to ISO 27001
EU AI Act (entered into force Aug 2024; obligations phased in from 2025 to 2027) Risk-based: prohibited practices, high-risk systems (strict obligations), transparency obligations (e.g. chatbots, generated content), general-purpose AI model obligations
Sector rules Model risk management in banking (e.g. SR 11-7 in the US), medical device rules for clinical AI

[!VERSION] EU AI Act timelines and guidance are being phased in and clarified; verify current obligations and dates with legal counsel.

Governance operating model

flowchart LR
  IDEA[AI use case proposal] --> REG[Register in AI inventory]
  REG --> RISK[Risk classification<br/>impact, data, autonomy, regulation]
  RISK -- low --> AUTO[Self-service with standard controls]
  RISK -- high --> REVIEW[Responsible AI review<br/>legal, security, privacy, ethics]
  AUTO & REVIEW --> BUILD[Build with platform guardrails]
  BUILD --> EVAL[Evaluation evidence<br/>quality, bias, safety]
  EVAL --> APPROVE[Approval gate]
  APPROVE --> RUN[Production with monitoring]
  RUN --> AUDIT[Periodic review / revalidation]

Key controls

  • AI inventory: purpose, owner, models, data, users, risk tier.
  • Documentation: model cards, system cards, data sheets, intended use and limitations.
  • Evaluation evidence before release and periodically: accuracy, robustness, bias/fairness, safety.
  • Human oversight proportional to risk; clear accountability.
  • Transparency: tell users when they interact with AI or receive AI-generated content where required.
  • Data governance: lawful basis, consent, IP/licensing of training data, retention.
  • Incident management for AI failures.

Interview questions and answers

How do you avoid governance slowing down AI adoption?

Tier by risk: most internal productivity uses get self-service through an approved platform with built-in guardrails; only high-risk uses (decisions about people, regulated domains, autonomous actions) go through full review. Automate evidence collection.