Japan's AI Act has no penalties and still creates a duty for users

The first thing to establish about Japan's AI Promotion Act is that it carries no penalties.

Act No. 53 of 2025, formally the Act on Promotion of Research, Development and Utilisation of AI Related Technologies, was promulgated on 4 June 2025 and came fully into force on 1 September 2025. Unlike the EU AI Act, its purpose is promotion rather than regulation.

With no penalties there is no concept of a violation. Many companies stop reading at that point. In practice a different reading is more useful.

Who Article 7 applies to

Article 7 is the provision worth knowing. It places an efforts duty not only on businesses that develop or supply AI, but on businesses that use it.

You do not need to be building models. Using AI in your operations is enough to be in scope. Given that 86.4% of Japanese companies use generative AI somewhere, that is close to everyone.

An efforts duty carries no punishment for falling short. It still matters in two situations.

When something goes wrong. If an AI output causes harm, the question becomes whether the company had appropriate arrangements in place. Having a statutory efforts duty on the books and no arrangements at all is a difficult position to explain.

When a customer asks. Large companies are beginning to include AI governance in procurement requirements. Being asked with reference to the Act and having no answer is a commercial disadvantage.

The relationship to the AI Business Guidelines

Version 1.2 of the AI Business Guidelines, updated in March 2026, clarified the treatment of AI agents and the human in the loop principle.

An agent is a configuration where the system executes multiple steps on its own. Because a person is not reviewing each step, the points where a human does intervene have to be decided in advance.

The guidelines are not law. As a reference for building the arrangements, they are practical.

What to actually have in place

There is no need to over prepare for a law with no penalties. But the following four are worth having regardless of the legislation.

1. An inventory of the AI services in use. Many companies do not know which department uses what. The audit takes less than a day.
2. A standard for what may be entered. Personal data, customer data, unpublished information. One page stating what is permitted.
3. A definition of where a human reviews output. Which process, who, and what they check. This matters most where agents are involved.
4. An escalation route. Who gets told, and who has the authority to stop it.

These four satisfy the efforts duty. They are also the skeleton of a working process design.

Read it as a premise, not a constraint

The AI Promotion Act signals that Japan intends to be a country that promotes AI rather than one that restricts it. For companies, the accurate reading is not that new constraints arrived, but that use is now officially assumed.

Once use is the assumption, not using it becomes the position that requires an explanation.