The First Legal Checklist a Company Launching an AI Service Should Build

Hello, I'm Noriaki Asato, Representative Attorney at LegalAgent.

New businesses built on generative AI are advancing rapidly in every field: adding features to existing SaaS products, deploying internal document-search AI and customer-support chatbots, and developing automatic image and text generation tools and AI agents that carry out work on a user's behalf. Amid this pace of development, a major practical challenge is how to balance the speed of building the product with confirming that it is legally safe.

If a company puts development speed first and postpones legal review on the assumption that "we can do it all right before release," a large amount of checking will be left just before launch. This is because an AI service simultaneously involves multiple issues, including personal information protection, copyright, the limitation of civil liability and AI governance. On top of that, it is difficult for the business to fully predict in advance what data users will input and what output the AI will return.

That is exactly why, for a company launching an AI service, the first useful tool is not a thick set of internal rules built from scratch, but a legal checklist that clearly sets out "what to keep checking at which stage of the business."

Where AI-Related Legislation and Government Guidelines Stand

In Japan, the Act on the Promotion of Research, Development and Utilization of Artificial Intelligence-Related Technologies, enacted in 2025, came fully into force in September of that year. Article 7 of this Act sets out, among other things, an obligation for utilizing businesses to make efforts toward active utilization and an obligation to cooperate with measures taken by the national and local governments. However, the existence of this promotion act alone does not provide a mechanism for determining the legality of an individual service in every respect. The Act on the Protection of Personal Information (APPI), the Copyright Act and the industry-specific regulatory laws must each be checked individually according to what the service actually does.

Meanwhile, public reference materials on the conduct of AI businesses and on civil liability have also been put in place, such as the AI Guidelines for Business (Japanese) issued by the Ministry of Internal Affairs and Communications and the Ministry of Economy, Trade and Industry (METI), and the Guide on the Interpretation and Application of Civil Liability in AI Utilization (Japanese) published by METI in April 2026. Although these differ from statutes themselves, they are materials referred to in practice when deciding what kind of management structure and accountability a business should establish. Application to an individual case is determined in light of the relevant laws and the specific circumstances.

The reality of AI legal work lies less in searching for an entirely new and unknown law than in the fact that introducing AI sets existing legal issues in motion all at once. Building a reasonable review procedure into service design and the development process is just as important as having knowledge of the statutory provisions.

Sorting Out Your Company's Position by Mode of Provision

When creating the checklist, the first task is to sort out your company's position. Identify whether your company falls under the role of AI developer, AI provider or AI user, and how that role differs from function to function. Depending on the role you play, individual items to be checked are added on top of the obligations that apply in common.

Are you developing and training your own foundation model, are you incorporating an external API model to provide an application, or are you merely using AI as an internal efficiency tool? Do you provide the AI's output directly to end users, or is it an AI agent that autonomously performs operations on behalf of users, such as placing orders with or sending data to external services? Where the service takes the form of an AI agent, practical controls need to be set, such as the scope of the permissions it can exercise, the situations in which prior human approval is required, the recording of execution logs and emergency stop procedures.

Even if you are simply building an application using an external provider's API, from the end user's perspective your company is the "provider of the AI service." If the data entered by users is sent to an external API, you must understand the external provider's terms of use, its data retention policy and whether the data is transferred to servers overseas. Another item to consider is making clear in your terms of use the extent to which you bear responsibility if the AI outputs an incorrect answer and causes a user loss.

On the other hand, even if your company uses AI only for internal operations, if it inputs customers' confidential information or personal data, consistency with nondisclosure agreements with business partners and with the APPI immediately comes into question. Even if you do not offer a service to outside parties, you cannot say the issue has nothing to do with you.

Risks Hidden in Input Data and What to Check in Advance

The data that users input is a matter to check at the early stage of service design. What wording will users enter into prompts, what files will they upload, and is there room for internal documents, unpublished contracts or trade secrets to be input? If you draft terms of use without confirming these realities, the design will diverge from how the service is actually used.

For input data, I recommend reviewing the following items as a checklist.

• Whether personal information may be input
• Concern that special care-required personal information or sensitive data may be mixed in
• The possibility that customer companies' trade secrets or information subject to confidentiality obligations may be mixed in
• The possibility that text, images, code or other material in which a third party holds copyright may be input
• Whether input data is sent to external AI providers or partners
• Whether there is a setting under which input data is used for additional training of an external model
• The retention period and storage arrangements for input prompts and log data
• Clear indication, in the terms of use and on the input screen, of information that must not be input
• For services for businesses, whether there are functions allowing an administrator to control and restrict employees' use

In B2B services for businesses in particular, situations can arise where employees of a customer company carelessly upload important internal materials. The more you increase the product's convenience, the more proper management of input data becomes the key to risk management.

This issue cannot be adequately addressed simply by writing a single line in the terms stating "personal information must not be input." If the screen's UI is designed to actively encourage file uploads, it is difficult to prevent such input through the wording of the terms alone. The design needs to combine on-screen warnings, input restrictions and contractual explanations.

Training Data, Generated Output and the Copyright Act

When building an AI service, you must also squarely address the rights in data used for machine learning and in data referenced in retrieval-augmented generation (RAG). Japan's Copyright Act contains provisions limiting rights in order to facilitate information analysis, but they do not lead to the conclusion that every use of data is permitted unconditionally. The Agency for Cultural Affairs' materials on AI and copyright (Japanese) also show the importance of legal analysis that clearly separates the AI development and training stage from the generation and use stage.

If you rely on Article 30-4 of the Copyright Act (Japanese) instead of a license for uses such as training, confirm that each requirement is met: the purpose is not to enjoy the thoughts or sentiments expressed in the work, the use is within the necessary extent, and the use does not unreasonably prejudice the interests of the copyright holder. If your company fine-tunes or additionally trains a model, check the sources from which the training data is collected and their terms of use, and if you obtain the data through web scraping or via an external API, make sure that this does not violate the terms of use of the target site. If you adopt a RAG configuration that searches internal documents or external materials to generate answers, you must carry out a rights check separate from training, namely whether you have lawful authority to reproduce, publicly transmit and display the source documents.

At the generation and use stage, two issues arise: technical and operational measures to limit the risk that the generated output resembles existing works, and how to allocate responsibility by contract where a user inputs a third party's work into a prompt. Although similarity of expression alone does not necessarily mean that infringement will immediately be found, copyright infringement can be established if similarity to the creative expression of an existing work and reliance on that work are recognized. Even then, you check whether a license or a rights-limitation provision applies. If you decide in advance on a contact point for receiving infringement claims, a process for removing problematic output and procedures for preventing recurrence, you can respond calmly when a claim actually comes in.

In the terms of AI services, one often sees statements such as "all rights in the output belong to the user." However, whether a generated work qualifies as a copyrightable work depends on the human's creative intent and the degree of creative contribution, and there is also concern that the output may contain expression that infringes a third party's rights. Even if the ownership of rights is set out in contract clauses, they have no effect to create a copyright that does not exist or to extinguish rights held by a third party. It is important to treat the clause on ownership of rights separately from the clause allocating responsibility between the business and the user.

Errors in Output and Managing Civil Liability

AI services are dogged by the problem of hallucination, in which the AI outputs plausible-sounding errors. In fields directly tied to users' significant decisions, such as law, medicine, accounting and finance, the losses arising from taking AI output at face value can be large. The duty of care and the degree of responsibility a business should bear also change depending on whether the AI merely provides supplementary reference information or is offered at a level that replaces human judgment.

METI's guide also sets out, taking into account the black-box nature and autonomous behavior of AI, how civil negligence liability and tort liability under current law should be understood. Rather than assuming that liability is avoided because AI is involved, the task is to consider system design, advance explanations and the operating structure in concrete terms.

The practical response begins with clarifying the position of AI output within the service. Decide whether the output is positioned strictly as a supplementary draft or search result, or whether users are to rely on it as the core of business decisions. Then decide at which business steps human review (human in the loop) will be built in, how to explain to users the limits of the output's accuracy, currency and legality, and what cautionary notices to display in situations requiring a professional's judgment. Operationally, prepare a flow for receiving inquiries when an incorrect answer is produced, for correcting or temporarily suspending the service, and for reporting, and decide on the necessary logs and retention periods, with due care for the handling of personal data and confidential information, so that the behavior at the time can be verified after the fact if a dispute arises.

I do not recommend a structure that tries to escape all legal liability through the disclaimer clause in the terms of use alone. Not only is there concern that the disclaimer may be invalidated by mandatory laws, starting with the Consumer Contract Act, but it does nothing to prevent the loss of corporate credibility when trouble actually occurs. It is important to manage the scope of the risks you accept realistically, in coordination with the service's screen design, warnings before use and the internal review flow.

Sending Personal Information Externally and Ongoing Review

Compliance with the APPI is also frequently discussed for AI services. Where users input personal information into prompts, you need to check points such as whether this is within the scope of your purposes of use; whether sending the data to an external AI provider or cloud server constitutes outsourcing or provision to a third party; whether it falls under provision to a third party in a foreign country; whether security control measures are in place; and whether you are prepared to report to the authorities and notify individuals in the event of a data breach.

In particular, when using an external API, the attitude of deciding that "we are just calling an external model, so it has nothing to do with us" will not hold. As long as you hold users' data and provide a service, how your company explains and manages the handling of that data needs to be confirmed.

In accordance with the Personal Information Protection Commission's General Guidelines (Japanese), review the specification of the purposes of use of personal information, security control measures for personal data and the supervision of outsourcees. Cases in which personal data is provided to an outsourcee located in a foreign country can also fall under Article 28 of the APPI, so confirm the legal basis, such as whether a statutory exclusion applies, whether there is a system for continuously taking equivalent measures, or whether the individual's consent has been obtained with prior provision of information. Simply turning off training use in the API settings does not immediately eliminate the characterization of the transaction as external provision of personal data.

When reviewing the privacy policy for an AI service, use the following items as the starting point.

• The items of personal information acquired and processed through providing the AI function
• A statement of specifically identified purposes of use
• The distinction between outsourcing and provision to third parties with respect to external AI providers, cloud infrastructure and analytics tools
• Whether data is transferred to third parties in foreign countries, and the legal basis
• The retention period and management procedures for prompts, log data and attachments
• Whether input data is used to train AI models
• The arrangements for receiving requests from individuals for disclosure, correction, suspension of use and the like
• The internal reporting and regulatory response flow in the event of a personal data breach or similar incident

With AI products, the destinations to which data is sent and the items acquired can easily change due to model changes or the addition of new features. You therefore need a mechanism for returning to the checklist each time a feature is added or an external API is changed, rather than relying on a one-time check at the initial release.

Designing Usage Rules Through the Terms of Use

In the terms of use for an AI service, even compared with ordinary SaaS, it is important to design clauses with users' concrete patterns of use in mind. Set out specific prohibited acts in advance, such as violations of law, infringement of rights, generating malware and uses contrary to public order and morals. Also specify, regarding input data and output, how far the business may use them within the scope of service improvement and maintenance, whether users are permitted to use the deliverables commercially, and how responsibility is to be allocated if a third party's rights are infringed.

For a service for businesses, make clear in advance the administrator's authority to review employees' usage history, measures for immediately suspending an account in the event of a violation of the terms, and the scope of the corporate customer's responsibility for misuse by its employees.

It is important to draft the terms of use not merely as a document listing your company's disclaimers, but as an agreement through which users and the business share an environment in which the service can be used with confidence.

Recording Decisions and Assigning Owners

The significance of an AI legal checklist is not limited to the legal function putting the brakes on development. It is a tool that allows management, the development team and the business team to discuss decisions on the same premises, such as whether a new feature may be released, whether an external API may be switched, or whether to have separate terms for business and individual users, and to keep a record of how those decisions were made. Keeping the grounds for decisions makes it easier to respond, with reasons, to future security checks from corporate customers and to inquiries from regulators.

At LegalAgent, attorneys well versed in generative AI and corporate law work together to provide legal support that goes beyond drafting contracts and terms of use, taking into account product feature design, data flows and internal operations. Information on setting up a structure in which outside attorneys are involved in product decisions in a way close to an in-house legal team is available on our Legal Outsourcing page.