AI Governance Category Formation Across Regulatory Jurisdictions

AI governance category formation across regulatory jurisdictions must contend with incompatible definitions. For platform and deep-tech companies, the language strategy starts with a stable description of what the technology does, which decisions it supports, and who owns its behavior. That description must then map to each jurisdiction’s categories. A persuasive product name cannot decide how the law classifies a system.

Why every compliance team is solving the wrong problem

Workday argued that it was a software provider. The employers using its technology had delegated traditional hiring functions without clearly assigning compliance responsibility. The court nevertheless allowed the case to proceed on an “agent” theory, under which Workday could be directly liable as an agent of those employers. That was not a final liability finding. But it exposed the gap between what a company calls its role and how a court can classify the functions it performs.

That gap is the problem a longer checklist cannot solve. Enterprise compliance often treats governance as an exercise in collecting rules, when the deeper difficulty is that jurisdictions disagree about what those rules should classify. A system can comply in one jurisdiction and fall short in another because the regimes disagree about whether it is regulated at all.

How the EU, UK, and US each answered the prior question differently

The EU’s AI Act organizes governance around the use case, classified by risk tier. Its four tiers are unacceptable, high-risk, limited-risk, and minimal-risk, with unacceptable uses prohibited. Classification centers on what the system does, rather than simply who built it or which industry bought it. Location offers no easy escape. The Act reaches providers, deployers, importers, and distributors whenever output is used in the EU, including a US provider serving European companies.

The UK instead bases governance on the sector, with the regulators already in place interpreting five cross-sector principles. Those principles are safety and robustness, transparency and explainability, fairness, accountability and governance, and contestability and redress. The same model can therefore receive different treatment in banking, healthcare, employment, or communications. The regulator’s domain changes the analysis. As of 2026, UK regulators are signaling less tolerance for opaque decisions involving creditworthiness, hiring, pricing, or access to essential services.

In the US, the unit of governance is itself contested as federal intent collides with state action. As of Q2 2026, nine states have active AI legislation, including CA, CO, IL, NY, TX, MA, WA, MN, and RI. “Algorithmic discrimination” has different legal definitions in New York, California, and Colorado. Federal intent does not erase those differences. The executive order does not itself preempt, suspend, or invalidate existing state laws, so further litigation, spending, or standard-setting would be required. A multinational therefore faces three different answers to “what counts as a regulated AI system,” not three versions of the same rulebook.

What the category systems share at the surface — and why that surface agreement is misleading

ISO 42001, published in December 2023, the NIST AI RMF, published in January 2023, and the OWASP Top 10 for LLMs from 2025 define enterprise benchmarks across jurisdictions. They are not legally mandatory in most places, but procurement teams and insurance underwriters use them to assess governance posture. This convergence has real commercial value. It lets organizations show their process controls and documentation without starting from scratch at every border.

The EU defines “high-risk” through enumerated uses in Annex III and regulated products in Annex I. The UK has no equivalent master list, while Colorado, California, and Texas draw their triggers at different points in the decision pipeline. A common transparency process cannot settle whether a system falls within a legal category in the first place.

Arcadia Impact’s 2026 mapping of twenty AI middle-power jurisdictions found that only 22 percent of the provisions were part of binding law. Three-quarters of the instruments that did bind did so without defining GPAI at all. Definitions clustered in guidance and strategy documents, where precision had the least legal consequence. The distinction is between form and force. Outside the EU, no jurisdiction in the mapped sample imposed a binding evaluation duty on a model developer, leaving enforcement dependent on demonstrated harm.

Why the EU's category architecture is the de facto global template — and why it hasn't been fully operationalized yet

The EU framework is the primary reference point for organizations with EU-facing operations because it combines a comprehensive horizontal law with extraterritorial reach. That makes it structurally unavoidable. Prohibited practices have been enforceable since February 2025, and GPAI obligations have applied since August 2025. The European Commission’s enforcement framework says that the enforcement powers of the AI Office and national authority apply from 2 August 2026, when some provisions become enforceable. Serious violations can attract penalties of up to 7% of global annual turnover, or a fixed maximum, whichever is higher.

CEN and CENELEC missed their original April 2025 deadline for harmonized technical standards and are targeting late 2026 at the earliest. The Digital Omnibus, Regulation (EU) 2026/1744, deferred high-risk deadlines accordingly. Standalone Annex III systems, including hiring and credit scoring, have a December 2, 2027 deadline, while high-risk systems embedded in Annex I products have an August 2, 2028 deadline. The delay addresses a practical dependency, because providers cannot demonstrate compliance through standards that do not yet exist.

Organizations are left with legal categories for which the harmonized basis for audits is still being written. It also makes version control a governance concern. The Commission proposed the amendments on November 19, 2025, and Parliament and Council reached political agreement on May 7, 2026, before adoption ahead of the original August 2026 deadline. A compliance plan based only on the original 2024 text would miscalculate every high-risk deadline by at least a year.

How agentic AI and distributed deployment have outpaced the category systems already in place

Existing categories assume discrete AI systems with traceable decision pipelines. Agentic systems act through many steps, tools, and data sources, which makes it harder to tell what unit is being governed. Gartner predicts that 40% of enterprise applications will include task-specific AI agents by the end of 2026. Each component may fall below a regulatory threshold even when the system as a whole makes high-stakes decisions.

This creates an incentive to design around category boundaries. The LLM governance literature identifies the possibility of firms restructuring pipelines so that individual components avoid classification while harm persists or moves elsewhere. Firms can do better at meeting the rules without making the same gains in safety. The loophole is the boundary itself. Under the current frameworks described across the three major jurisdictions, that outcome would be discovered only after implementation.

LAAF’s systematic review also finds that governance work is hard to measure. Human expert reviewers often disagreed about what the documents covered because they used different terms for sectors and risk domains. If people cannot consistently identify what a governance document covers, that inconsistency can cascade into audit failure. Shadow AI brings further risk, as research finds that average breach costs are much higher in organizations with high shadow AI than in those with low or none. Agentic AI stretches the boundaries of the regulated system, while shadow AI makes the system harder to see at all.

Why liability doesn't concentrate when ownership is distributed without structure

When vendors, deployers, and internal teams have no clear structure for accountability, more than one party can be exposed to liability. The Workday case illustrates why a software-provider label is not enough to resolve that allocation. Allocating liability requires examining the function performed and the responsibility attached to it. An organizational chart with several arrows is not necessarily an answer.

The LAAF review identifies four persistent gaps, namely under-specified human oversight, missing shared accountability metrics, disciplinary disconnection, and limited empirical evaluation. These are organizational weaknesses. Accountability links an actor to a forum, such as a regulator, court, or auditor. If the actor and forum classify the system differently, even good-faith explanations may fail to establish who was answerable for its behavior.

Checks by buyers and insurers make this matter now, even where rules are only partly enforced. Fortune 500 RFPs call for ISO 42001 certification, SOC 2 with its existing criteria applied to AI risks, or third-party bias audits before a contract is awarded. Cyber and professional liability insurers are excluding or conditioning coverage for unaudited AI systems. Underwriters ask about governance policies, bias testing, and algorithm impact assessments. Ambiguous ownership can therefore obstruct a contract before it becomes a courtroom problem.

The organizational response is structural clarity about roles and descriptions. Terms such as provider, deployer, agent, and operator have different legal weight across jurisdictions, so internal language needs a canonical account of who owns what. This account is the basis for assigning responsibility consistently to employees, vendors, and counterparties, while recognizing that each jurisdiction will apply its own legal classification.

What the geopolitical dimension means for organizations that hoped interoperability would arrive

Shared standards already let organizations show how they handle governance across borders, and framework-mapping can show which controls they have in common. That work reduces duplication without requiring governments to agree on everything. It is necessary infrastructure for multinational compliance. The mistake is expecting a controls map to decide which systems are regulated or who holds liability when they cause harm.

China keeps stronger controls directed by the state, while Gulf states use frameworks that put innovation first to draw investment. Singapore, Japan, and African jurisdictions are building approaches shaped by domestic industrial policy. Different rules can be the intended outcome. AI governance is also an instrument of geopolitical competition, so alignment cannot be treated as a delayed software update.

The Arcadia mapping adds an enforcement dimension to that divergence. Almost every evaluation body in its sample was constituted without power to act on its findings. A body may therefore exist on paper without being able to enforce rules as an enterprise expects. At the more fractured end, sovereign regimes could force companies to create separate products or leave particular markets. That is a live planning scenario, and organizations cannot base their operating model on near-term convergence.

The strategic literacy organizations actually need to operate across these borders

Organizations can start with a unified control architecture built from the shared standards. Above it, organizations need teams that can work together to define each AI system, decide who owns its behavior, and map those facts to each jurisdiction’s categories. Compliance depends on teams working together to establish a stable account of what the system actually does.

For a platform technology that buyers or analysts keep putting in the wrong bucket, the naming discipline starts with function and responsibility. The internal description needs to state what the system does, which decision it makes or supports, who acts as provider or deployer, and which annex or trigger applies in each jurisdiction. A new category name is not a substitute. The same underlying account must support different regulatory interpretations without becoming three incompatible stories about the product.

That is where narrative infrastructure becomes a governance function. Storied’s Narrative OS is designed to support the canonical language layer that makes descriptions durable across teams and contexts. In this setting, its relevant purpose is not a sharper slogan but a governed basis for compliance claims. Clear language must reflect an actual assignment of accountability. Otherwise, a polished description only makes the ambiguity easier to circulate.

The EU’s revised deadlines show why this capability cannot be a one-time project. CEN and CENELEC have not yet delivered the technical standards on which the high-risk classifications depend, Colorado has replaced one AI statute with another, and the US federal-state contest has no near-term resolution. The categories themselves will keep moving. Organizations need infrastructure that keeps track of those changes, updates the descriptions they use within their teams, and keeps their claims up to date across jurisdictions at the same time.