EU AI Act vs Japan AI Promotion Act: Binding Rules vs Voluntary Guidance for Global Enterprises

Two instruments with different legal natures

Regulation (EU) 2024/1689 sets binding rules built on a risk-based approach. It applies to providers that place AI systems on the market or put them into service in the Union, or that place general-purpose AI models on the market, whether the provider is established in the Union or in a third country. A system operated from outside the EU can therefore be in scope. The rules impose obligations on both providers and deployers.

Japan's AI Promotion Act is framed differently. It is oriented toward promoting AI and relies on voluntary guidance and government coordination rather than a tiered set of binding obligations. For a compliance team this is the central distinction. One instrument tells you what you must do and attaches penalties. The other shapes expectations of conduct that an enterprise chooses how to meet.

The EU timeline is already moving. As of November 2026, the governance structure, penalty provisions and obligations for providers of general-purpose AI models have applied since 2 August 2025. The Digital Omnibus on AI entered into force on 27 July 2026 and moved the high-risk application dates from 2 August 2026 to 2 December 2027 for stand-alone high-risk systems and 2 August 2028 for high-risk systems embedded in products. The Commission cited delayed availability of standards. Treat dates as subject to further change.

Side-by-side comparison

The table summarises the dimensions that most affect operating requirements. It describes the structural contrast, not a clause-by-clause mapping.

• Legal force: EU: binding regulation. Japan: promotion-oriented framework with voluntary guidance.
• Scope: EU: providers placing systems or general-purpose models on the EU market, including from third countries. Japan: assess scope against each system's Japanese footprint.
• Structure of obligations: EU: risk-tiered, differing by category, such as high-risk systems, general-purpose models and AI that interacts with humans, and by role as provider or deployer. Japan: guidance-led expectations rather than tiered duties.
• Enforcement: EU: governance structure including an AI board of member-state representatives, with penalties based on a share of prior-year global turnover or a set amount, whichever is higher. Japan: government coordination and voluntary guidance.
• Evidence expectations: EU: a documented, maintained risk management system for high-risk systems and checkable duties such as AI disclosure. Japan: enterprise-determined, so evidence practice is a design choice.
• Change over time: EU: amendable through delegated and implementing acts, including reviews of the Annex III high-risk list. Japan: guidance can be updated as government coordination develops.

What binding EU obligations mean at runtime

Several EU requirements imply controls that operate while systems run, not only before deployment. The Act requires a risk management system for high-risk AI systems to be established, implemented, documented and maintained. That implies a durable record of risk controls over the life of the system, not a one-time design assessment. Conversational AI that could be mistaken for a person must inform the human that they are interacting with AI. This is a user-facing duty that can be checked in operation.

Obligations also vary by role and tier. Providers of general-purpose models above a capability threshold carry additional risk management and cybersecurity duties, and these are provider-side obligations, not automatically deployer duties. A deployer using a model supplied by someone else has a different position from the provider that placed it on the market. Policy configuration should therefore be keyed to each system's role and classification, not applied as a single global setting.

The operational consequence is that the evidence a regulator may want is evidence of behaviour: what the system was permitted to do, what it did, and when a human was brought in. Controls that act in the execution path, such as allow, redact, block or escalate for human approval, can generate that evidence as a by-product of enforcement. Any mapping from specific controls to specific articles needs legal review.

Uniform baseline or jurisdiction-specific configuration

A single baseline built on the stricter, binding regime has clear advantages. It reduces the number of policy variants to maintain, avoids inconsistent behaviour for the same system across regions, and means evidence is produced in one consistent form. Because Japan's framework relies on voluntary guidance, a control set designed for binding EU obligations is a reasonable starting point. Counsel should confirm how it interacts with Japanese guidance before it is treated as sufficient.

A uniform baseline has costs. Controls designed for high-risk obligations may add friction to low-risk uses that the EU regime does not tier the same way. For that reason the baseline should be uniform only for controls that are cheap to apply everywhere, such as disclosure checks, escalation paths and evidence generation. Role- and tier-specific controls belong in configuration.

Four criteria help decide where to split by jurisdiction. First, legal force: does a binding duty exist for this system, given its role and risk tier? Second, data residency: where must interaction data and evidence be stored? Third, retention: how long must evidence be kept, and does that differ by region? Fourth, audit audience: who will ask to see evidence, and in what form? Where the answers differ, configure by jurisdiction. Where they match, apply one policy.

Where runtime enforcement and evidence fit

The comparison points to one practical requirement: a way to enforce policy at the moment an AI system acts, and to retain evidence of the result. Trussed addresses that layer. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants. Actions include allow, redact, block, rate limit and escalate for human approval, applied to individual interactions and tool calls. The enforcement produces continuous, audit-ready evidence mapped to the EU AI Act, among other frameworks.

Trussed deploys self-managed in the customer cloud, with configurable retention and data residency, which supports the jurisdiction-specific configuration described above. It is a runtime control layer for regulated enterprises, not a replacement for classification, legal interpretation or risk management processes. Those remain the compliance team's work.