Skip to content
Frameworks

What applies where and from when.

We set out the instruments we work against provision by provision. Where a date has moved we say so. A governance programme built on a superseded timeline is worse than none.

Coverage

Maintained
Jurisdictions
4
Provisions tracked
27
In force today
17
Still upcoming
10
01European Union

The AI Act as amended by the Digital Omnibus.

The deferral of the high-risk regimes bought time without reducing the work. Transparency obligations are live now. Conformity evidence takes eighteen months to assemble and the technical file has to describe a system that already exists.

3 in force5 upcoming
Primary sourceRegulation (EU) 2024/1689 on EUR-Lex(opens in a new tab)
  • Art. 5In force

    Prohibited practices

    Eight categories have applied since February 2025. The Omnibus adds two more from 2 December 2026. They cover non-consensual intimate imagery of an identifiable person and child sexual abuse material. The prohibitions are absolute and have no conformity route.

    1.8 years ago
  • Art. 50In force

    Transparency for certain systems

    Interaction disclosure, synthetic-content marking, deepfake disclosure and notice of emotion recognition and biometric categorisation. It applies regardless of risk classification. That is why it catches teams that assumed they were out of scope. Generative systems already on the market before 2 August 2026 have until 2 December 2026 to add machine-readable marking.

    4 months ago
  • Ch. VIn force

    General-purpose AI models

    Technical documentation for downstream providers, a public training-content summary and a copyright policy. Models meeting the systemic-risk threshold add adversarial evaluation, incident reporting and cybersecurity duties.

    1.3 years ago
  • Art. 6 · Annex IIIUpcoming

    High-risk classification

    Annex III lists eight areas: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration and justice. The Art. 6(3) filter is for systems that perform only narrow procedural tasks. Classification is a written determination you have to be able to defend.

    1 year from now
  • Art. 9–15Upcoming

    Requirements for high-risk systems

    A continuous risk management system; data and data-governance obligations; logging; instructions for use; human oversight design; accuracy, robustness and cybersecurity commensurate with intended purpose.

    1 year from now
  • Art. 11 · Annex IVUpcoming

    Technical documentation

    Nine sections: system description, development process, monitoring, functioning and control, performance metrics, risk management, lifecycle changes, standards applied, the EU declaration of conformity and the post-market monitoring plan. Retained for ten years after placing on the market.

    1 year from now
  • Art. 43 · Annex VI–VIIUpcoming

    Conformity assessment

    Annex III points 2 to 8 follow internal control under Annex VI with no notified body. Biometric systems under point 1 go to a notified body under Annex VII where harmonised standards were not applied in full. Where they were the provider may choose either route. SMEs and small mid-caps may submit the Annex IV information in a simplified form.

    1 year from now
  • Art. 72Upcoming

    Post-market monitoring

    A documented plan proportionate to the system. It collects performance data across the system's lifetime and feeds both the risk management system and the serious-incident reporting duty under Article 73.

    1 year from now
02Canada

Canada has no AI act. Most people get this wrong.

AIDA died with Bill C-27 on prorogation in early 2025 and has not been reintroduced. Canada now regulates AI through existing instruments that already bind you. Two bills tabled in June 2026 would add AI-related duties. Neither is law yet. Building a programme around a statute that was never passed is the most common Canadian governance error we are asked to correct.

5 in force3 upcoming
Primary sourceOSFI Guideline E-23 — Model Risk Management (2027)(opens in a new tab)
  • National AI StrategyIn force

    AI for All

    Released 4 June 2026 as a five-year plan. The strategy sets direction, funding and institutional capacity. It imposes no binding obligations. The AI duties it promised have arrived as bills instead of a standalone AI statute. Bill C-36 covers privacy and Bill C-34 covers online safety.

    6 months ago
  • PIPEDAIn force

    Federal private-sector privacy

    Governs the personal information flowing through training data, inference and logging: accountability, limiting collection, accuracy, safeguards and access rights that extend to the personal information held inside a model pipeline.

  • Law 25, s. 12.1In force

    Québec automated decisions

    Where a decision is based exclusively on automated processing, the individual must be informed at the time of the decision and, on request, told the personal information used, the principal factors and parameters, and given the right to submit observations to a human who can review it. A generic privacy notice does not discharge this.

    3.2 years ago
  • OSFI E-23 (2027)Upcoming

    Model risk management

    Effective 1 May 2027 for federally regulated financial institutions including insurers and branches. It sets out risk-based governance across the model lifecycle: identification, development, validation, deployment, monitoring and decommissioning. AI/ML models and their data dependencies get explicit guidance.

    5 months from now
  • AI Directive · EDSTA, 2024In force

    Ontario public sector

    Ontario's Responsible Use of Artificial Intelligence Directive has bound ministries and provincial agencies since 1 December 2024. The Enhancing Digital Security and Trust Act lets the province extend AI duties to other public-sector entities by regulation. As of November 2026 no regulation has done so. Vendors meet these rules through contract terms and not through the Act itself.

    2 years ago
  • PHIPA · provincial healthIn force

    Health information custodians

    Clinical and administrative AI that touches personal health information sits under provincial custodianship rules. Their limits on use, disclosure and de-identification constrain both training and vendor arrangements.

  • Bill C-36Upcoming

    Protecting Privacy and Consumer Data Act

    First reading 15 June 2026. It would replace the privacy part of PIPEDA. Section 63 requires an explanation on request of an automated decision with a significant effect on a person. The explanation covers the personal information used and the principal factors. The person can then make representations to someone able to review the decision. It is not yet law.

  • Bill C-34Upcoming

    Safe Social Media Act

    First reading 10 June 2026. It passed second reading on 29 October 2026 and is in committee. It would enact a Digital Safety Act that covers AI chatbot services. Regulated operators would have to limit harmful content, follow self-harm protocols and stop a chatbot posing as a human or a licensed professional. It is not yet law.

03United States

A framework and a patchwork.

There is no federal AI statute. Buyers ask for a NIST AI RMF mapping because it is the common vocabulary in enterprise procurement and federal contracting. After that they ask about whichever state regimes touch your users.

3 in force2 upcoming
Primary sourceNIST AI Risk Management Framework(opens in a new tab)
  • NIST AI 100-1In force

    AI Risk Management Framework 1.0

    Four functions (Govern, Map, Measure and Manage) with categories and subcategories. It is voluntary. That is why it has become the de facto contractual baseline: a customer can ask you to map to it without imposing a foreign statute.

  • NIST AI 600-1In force

    Generative AI Profile

    Twelve risks that generative AI creates or makes worse. They include harmful bias and homogenization, confabulation, data privacy and information integrity. Suggested actions are organised under all four RMF functions. For LLM-based products this is the practical starting point.

  • NIST IR 8596Upcoming

    Cyber AI Profile

    A preliminary draft since December 2025. It bridges AI risk management with Cybersecurity Framework 2.0. It is worth tracking now because it is where AI security controls and existing CSF programmes will converge.

  • SB 26-189Upcoming

    Colorado

    Effective 1 January 2027. The replacement law drops the algorithmic-discrimination duty of care and deployer impact assessments. Developers must give deployers documentation. Deployers must give notice before a consequential automated decision and an explanation after an adverse one. Consumers can correct inaccurate data and ask for human review. The Attorney General's rules were still in draft after a hearing on 26 October 2026.

    1 month from now
  • State patchworkIn force

    Sectoral and state rules

    Automated employment decision auditing, insurance model governance bulletins, health AI disclosure rules and comprehensive state privacy laws with profiling opt-outs. Scope is determined by where your users are. Where you are based does not decide it.

04Standards

What certification is for.

Under the AI Act only a harmonised standard cited in the Official Journal gives a presumption of conformity. The first was cited in October 2026. A certifiable standard such as ISO/IEC 42001 does something else. It gives you a management system an auditor can test and a shared structure a buyer can accept in place of a bespoke questionnaire. That is usually the commercial reason to do it.

6 in force
Primary sourceISO/IEC 42001:2023(opens in a new tab)
  • ISO/IEC 42001:2023In force

    AI management system

    The certifiable AIMS standard. It covers context, leadership, planning, support, operation, evaluation and improvement. Annex A holds the controls and Annex B the guidance. Adopted in Europe as EN ISO/IEC 42001:2026. It is not cited in the Official Journal under the AI Act. On its own it gives no presumption of conformity.

  • EN 18286:2026In force

    Quality management system for the AI Act

    The first harmonised standard cited under the AI Act. Its reference was published in the Official Journal on 14 October 2026. A provider that follows it is presumed to meet the quality management requirements it covers under Article 17.

    1 month ago
  • ISO/IEC 42005:2025In force

    AI system impact assessment

    Published May 2025. It supplies a method for the impact assessment that 42001 requires. The method covers scoping, affected parties, documentation and analysis of benefits and harms across the system lifecycle.

  • ISO/IEC 42006:2025In force

    Requirements for certification bodies

    Additional requirements to ISO/IEC 17021-1 for bodies auditing AI management systems. It matters because it determines auditor competence and audit duration. Those shape the certification timeline.

  • ISO/IEC 23894:2023In force

    AI risk management guidance

    ISO 31000 applied to AI. Useful where an enterprise risk function already runs on 31000 and wants AI risk expressed in the same language rather than a parallel taxonomy.

  • SOC 2In force

    Adjacent but not equivalent

    A SOC 2 report covers the service organisation's controls. It does not cover the behaviour of the model. Buyers increasingly ask for both. Presenting one as the other is a fast way to fail a security review.

Sector overlays

Where AI duties land on a regime you already have.

In regulated sectors the AI obligations arrive on top of an existing framework. The interaction between the two usually sets the timeline.

4 overlays
  • Health & life sciences

    Clinical decision support crosses into a regulated medical device faster than most teams expect. In the EU that usually makes it high-risk through Annex I and the Medical Device Regulation. Annex III reaches emergency triage, eligibility for public health benefits and health insurance pricing. Provincial custodianship rules and the software-as-a-medical-device pathways at Health Canada and the FDA can attach to the same feature.

    • SaMD
    • PHIPA
    • Annex I (MDR)
    • Annex III §5
  • Financial services

    The operative regime is model risk management rather than AI policy. E-23 lifecycle controls, validation independence and the interaction with consumer-protection and fair-lending expectations govern what you can ship and how it must be evidenced.

    • OSFI E-23
    • Annex III §5(b)
  • Employment & HR

    Screening, ranking and evaluation tools are high-risk under Annex III and separately regulated in several US jurisdictions. Human oversight design and bias testing are the two areas that fail review most often.

    • Annex III §4
    • NYC LL 144
  • Public sector & procurement

    Ontario's AI directive binds ministries and provincial agencies. Its requirements reach vendors through contract terms. Directive-driven algorithmic impact assessment appears in federal and provincial procurement. The answer is usually needed before award.

    • Ontario AI Directive
    • AIA (TBS)

Note on this page

This summary is maintained for orientation. It is not legal advice and does not substitute for the instruments themselves. Applicability turns on facts specific to your system, your role in the value chain and the markets you place it on. Where a determination matters we write a reasoned one against your system instead of a general note.

Applicability

Most scoping questions are answered in a call.

Bring the system and the markets. We will tell you which of the above attaches and which does not.