Enterprise AI Governance Frameworks: Risk Tiers, Model Documentation, NIST AI RMF, ISO 42001 and the EU AI Act, Colorado and California Deadlines
How large organisations structure AI governance: tiered risk classification, accountable owners, model cards and documentation, vendor and API controls, the frameworks they use (NIST AI RMF, ISO/IEC 42001) and the key regulatory dates under the EU AI Act, Colorado's AI Act and California's ADMT rules.
AI governance has become an operational requirement rather than a policy statement. EU AI Act obligations are phasing in, Colorado’s AI Act took effect on June 30, 2026, and California’s rules on automated decision-making technology are rolling out. The programmes that work share a common shape: an inventory of AI systems, risk tiers that put light controls on low-risk uses and deep review on high-impact decisions, named owners, documentation such as model cards, and contractual controls for third-party models. This guide sets out that structure, the frameworks companies use and the regulatory dates to plan around.
Key regulatory dates
| Regime | Milestone | Date |
|---|---|---|
| EU AI Act | Entered into force | August 1, 2024 |
| EU AI Act | Prohibited practices and AI literacy obligations apply | February 2, 2025 |
| EU AI Act | General-purpose AI model obligations apply | August 2, 2025 |
| EU AI Act | Most remaining provisions apply, including transparency duties | August 2, 2026 |
| EU AI Act | High-risk rules for Annex III systems (such as hiring and credit) | December 2, 2027, per the amended timeline |
| EU AI Act | High-risk rules for AI in products covered by Annex I | August 2, 2028, per the amended timeline |
| Colorado AI Act | Enforcement begins (delayed from February 1, 2026) | June 30, 2026 |
| California CCPA regulations | New rules covering risk assessments and ADMT take effect, with phased compliance | From January 1, 2026 |
Sources: European Commission; AI Act implementation timeline; Wikipedia on the Colorado AI Act; California Privacy Protection Agency. The EU timeline for high-risk systems was pushed back after the Commission’s late-2025 simplification proposals, so check the current dates before planning.
The penalties are significant: under the EU AI Act, violations of the prohibited-practices rules can draw fines of up to €35 million or 7% of worldwide annual turnover, whichever is higher (Article 99).
A three-tier risk model
Many enterprises use a tiered classification that maps roughly to the EU AI Act’s categories but adds practical detail:
| Tier | Examples | Required controls |
|---|---|---|
| Prohibited | Social scoring; emotion recognition in the workplace; manipulative techniques | Not allowed; screen new projects against the list |
| High risk | Hiring and promotion decisions, credit and insurance underwriting, patient triage, access to essential services | Documented human oversight, bias and performance testing before launch, monitoring, incident processes, impact assessments where required |
| Lower risk | Internal drafting assistants, summarisation with human review, analytics, customer chatbots with clear disclosure | Standard software review for security, privacy and access; usage guidelines; transparency where required |
The classification should be by use case, not by tool. The same language model can be lower risk when drafting internal emails and high risk when screening job applicants.
Accountability and roles
Effective programmes name an owner for each AI system who is accountable for its performance and compliance, and a central governance function that maintains the inventory, reviews high-risk classifications and resolves disputes. Shadow AI is a common blind spot here; see our guide to shadow AI governance. Many organisations place that function under the chief risk officer, general counsel or chief data officer rather than inside the teams building models, so that the people building a system are not the only ones deciding whether it is safe to deploy. A cross-functional review group, drawing on legal, security, privacy, HR and the business, handles high-risk approvals.
Documentation: model cards and beyond
Model cards, proposed by Mitchell et al. in 2019 (arXiv:1810.03993), document a model’s intended use, training data characteristics, evaluation results, performance across groups and known limitations. They are now common practice, and Hugging Face uses them for models on its hub. For enterprises, documentation serves several purposes at once: it supports regulatory requirements such as EU AI Act technical documentation, it records what conditions a model was validated under when it moves between teams, and it gives auditors and incident responders a starting point.
Useful records for each system include data sources and lineage, evaluation and bias-testing results, human oversight arrangements, change history and an incident log.
Governing third-party models and APIs
Most enterprises do not train foundation models; they use APIs or AI features built into SaaS products. Because these models cannot be inspected directly, governance relies on:
- Contracts: permitted uses, data handling, no training on customer data, retention, audit rights and incident notification.
- Vendor assessments: security certifications, published model documentation and evaluation results.
- Your own testing: evaluating the model on your use case before deployment and after vendor updates.
- Usage controls: restricting which data classes may be sent and logging usage.
The frameworks enterprises use
- NIST AI Risk Management Framework 1.0 (January 2023): voluntary guidance organised around four functions, Govern, Map, Measure and Manage, with a Generative AI Profile added in July 2024 (NIST AI RMF).
- ISO/IEC 42001:2023: the first certifiable management-system standard for AI, comparable in structure to ISO/IEC 27001 for information security. Certification can help demonstrate governance maturity to customers and regulators.
- EU AI Act: sets binding obligations by risk level for providers and deployers placing AI on the EU market or affecting people in the EU.
Still unresolved
- Fine-tuned open models: Responsibility when a company fine-tunes an open-weight model and deploys it can shift toward the company as provider under the EU AI Act, while U.S. law remains less settled.
- State-by-state fragmentation in the U.S.: Colorado, California and other states take different approaches, and federal preemption has been debated but not enacted.
- Agentic systems: AI agents that take actions across systems raise new questions about permissions, monitoring and accountability that most frameworks address only partly.
A minimum viable governance programme
- Inventory every AI system and vendor feature in use, including AI features built into SaaS tools.
- Classify risk by use case, for example internal drafting versus hiring or credit decisions.
- Assign owners accountable for each system’s performance and compliance.
- Document and test: data sources, evaluation results, bias testing where relevant and incident logs.
- Review vendor contracts for data use, training opt-outs and liability.
- Train staff on AI literacy and on which data must never be entered into AI tools.
Our analysis
Governance programmes fail when they become a paper exercise run by one committee. The effective ones put light controls on low-risk uses and save in-depth review for high-impact decisions, which keeps approval fast enough that employees do not route around it. With EU high-risk deadlines moved to late 2027 and 2028, organisations have more time, but not a reason to wait: building inventories, documentation and vendor controls now is far cheaper than retrofitting them under deadline pressure.