SAKURA Law Office | Legal Update by Managing Partner Kenshiro Michishita
September 24, 2026
SAKURA Law Office Managing Partner Kenshiro Michishita
Executive Summary
A corporate generative AI policy should not be drafted as a catalogue of prohibitions. Its function is to define the conditions under which the company can use AI safely, lawfully and effectively. An effective framework should identify approved AI environments, connect information classification to permitted uses, distinguish low-risk from high-risk use cases, define when meaningful human review is required, allocate responsibility for approvals and incidents, address shadow AI and AI agents, and align internal rules with the actual contractual and technical settings of each AI service. Japanese law does not, as of September 24, 2026, require every company to adopt a document with the specific title “Generative AI Use Policy.” Nevertheless, where employees use AI in ordinary business operations, a clear internal governance framework can be critical to operationalizing obligations relating to personal information, trade secrets, copyright, contractual confidentiality, cybersecurity, employment and accountability. The objective is not to create a perfect policy on day one. It is to establish enforceable minimum rules, provide employees with a safe path for legitimate AI use, and continuously improve the framework as technology, use cases, contracts and regulation evolve.
SAKURA Law Office is pleased to publish the sixth installment of the Legal Update series by Managing Partner Kenshiro Michishita, entitled “How to Draft a Corporate Generative AI Policy and Internal Guidelines — Internal Rules and Operating Practices for Effective AI Governance (2026).”
The use of generative AI within companies has moved beyond pilot projects and isolated experimentation. Generative AI is increasingly embedded in drafting, translation, research, software development, contract review, customer support, internal search, marketing, recruitment, human-resources functions, management analysis and other day-to-day operations. More recently, AI agents capable of connecting to email, cloud storage, customer relationship management systems, internal databases and other business systems are beginning to perform actions rather than merely generate text.
As AI use expands, so do the legal and operational issues associated with personal information, trade secrets, third-party confidential information, copyright and other intellectual-property rights, inaccurate output, bias, information security, employment management, external communications and contractual liability. Even where management has formally approved AI adoption, a governance framework will remain nominal if the company cannot determine who is using which AI system, what information is being submitted, and how AI-generated outputs are being used in actual business decisions.
The appropriate response is neither an abstract statement of principles nor a blanket prohibition on generative AI. Companies need operating rules that employees can apply in practice: which AI may be used, by whom, for what purposes, with what categories of information; which use cases require prior approval; how AI output must be reviewed; and how suspected incidents must be reported, contained and investigated.
Japan’s December 2025 Guidelines for Ensuring the Appropriate Research, Development and Utilization of Artificial Intelligence-Related Technologies, adopted under the AI Act, emphasize a risk-based approach, active stakeholder involvement, end-to-end AI governance and agile responses. In March 2026, the AI Guidelines for Business were updated to Version 1.2. In July 2026, the Information-technology Promotion Agency of Japan (IPA) published practical guidance addressing enterprise-wide AI governance, internal AI-use guidelines, shadow AI, the AI-system lifecycle and risk assessment for generative AI and AI agents.
Against that background, this Legal Update explains, as of September 24, 2026, how companies should design a corporate generative AI policy and internal guidelines not merely as template language, but as an operating system for effective AI governance.
1. A Generative AI Policy Is Not Merely a Document Deciding Whether AI May Be Used
A common drafting mistake is to produce a policy that lists AI risks and prohibited conduct without explaining how legitimate use may proceed. A document composed almost entirely of prohibitions may appear conservative from a legal perspective, but employees may still be unable to determine what is permitted. The result may be reduced productive use or, more seriously, informal AI use outside the company’s approved channels.
An effective policy should therefore define not only prohibited conduct, but also permitted use, use requiring approval, categories of information that may not be entered, review requirements for AI output, responsible functions, incident-reporting obligations and procedures for updating the framework. Properly designed, an AI policy is less a restriction document than a set of conditions under which the company affirmatively permits AI to be used.
Companies also differ substantially in AI maturity. A company that has already deployed enterprise AI across multiple functions requires a different level of detail from a company beginning a limited proof of concept. Copying another company’s policy, or an online template, without first understanding actual internal use is unlikely to produce effective governance.
2. Japanese Law Does Not Require Every Company to Adopt a Document with the Specific Title “Generative AI Use Policy”
As of September 24, 2026, there is no Japanese statute that requires every company, in a uniform form, to adopt an internal rule specifically titled a “Generative AI Use Policy.” The absence of such a document should therefore not be treated, by itself, as an automatic violation of law.
That does not mean internal governance is optional in practice. Depending on the use case, Japan’s Act on the Protection of Personal Information, Unfair Competition Prevention Act, Copyright Act, contractual confidentiality obligations, employment laws, sector-specific regulation and other existing legal regimes may apply. The AI Act-related guidelines and the AI Guidelines for Business also emphasize risk identification, risk assessment, controls proportionate to potential impact and continuous improvement of AI governance.
Accordingly, where employees use generative AI in ordinary business operations, existing information-security policies, privacy policies, confidential-information rules and employment regulations may not adequately address AI-specific scenarios. Even where no formal statute mandates a document with a particular name, written AI rules can be important to operationalize corporate duties of care, information management, contractual compliance, quality control and accountability.
3. Before Drafting the Policy, Map How AI Is Actually Being Used
The drafting process should not begin with wording. It should begin with discovery. The company should first identify, as far as reasonably possible, how AI is actually being used across the organization.
This should include not only officially procured generative AI services, but also services accessed through personal accounts, AI functions embedded in office software and web-conferencing tools, browser extensions, coding assistants, translation tools, image generators, AI embedded in CRM systems, and AI used by contractors or other service providers in performing work for the company.
Knowing the name of the tool is not enough. The company should understand which function uses it, for what business purpose, what information is submitted, how output is used, and what systems the AI can access. A policy written without that factual map risks diverging from actual workflows and may unintentionally increase shadow AI.
4. A Three-Layer Structure Often Works Better: AI Principles, AI Use Policy and Operating Standards
If every technical detail is placed into a single policy, the document may require formal amendment whenever a model name, feature, setting or service provider changes. In practice, it is often more workable to separate rules according to how frequently they are expected to change.
The first layer may be an “AI Principles” or “AI Governance Policy” setting out the company’s high-level approach to AI, including human-centric use, safety, legal compliance and accountability. The second layer may be an “AI Use Policy” establishing employee obligations, approval procedures, prohibited conduct and incident reporting. The third layer may consist of operating standards, user manuals, approved-services registers, information-classification matrices and specific approval workflows.
Approved service names, plans and settings change frequently. Keeping those details in an appendix or operating standard that can be updated by the responsible function allows the company to preserve stability at the policy level while responding more agilely to technological change.
5. Define Scope by Function, Not Merely by the Label “Generative AI”
If a policy applies only to “ChatGPT and other generative AI services,” AI features embedded in ordinary SaaS products and business systems may fall outside its wording. AI capabilities are now routinely integrated into office suites, meeting systems, CRM platforms, cloud storage, development environments and search tools.
The policy should therefore consider whether its scope includes not only standalone generative AI services, but also AI models, embedded AI features, AI agents and other functionality that automatically generates text, images, audio, code, recommendations, decisions or other processing based on input information.
The company should also consider personal devices and personal accounts. From an information-governance perspective, it may be more appropriate to define scope by whether AI is being used in connection with company business or company information than by whether the user is operating a company-issued device.
6. Allocate AI Governance Responsibility Without Making Legal the Sole Owner
An effective policy should identify who administers the framework, who approves AI services, who reviews high-risk use cases and who leads incident response.
AI governance cannot be completed by the legal department alone. Information security, privacy, HR, internal audit, corporate planning, IT, procurement and business teams may all need to participate. In a larger organization, requiring the legal team to pre-approve every AI use can become operationally impossible and may encourage employees to bypass the formal process.
The better design is often to provide clear criteria that allow ordinary, low-risk use to proceed at the business level, while escalating higher-risk use cases to a specialist function or AI governance committee. Senior management need not approve individual prompts, but it should exercise appropriate oversight over which material business processes use AI and which categories of risk the company is willing to accept.
7. Establish an Approved-AI Framework Rather Than Trying to Solve Shadow AI Through Prohibition Alone
IPA’s July 2026 practical guidance on generative AI and AI agents identifies enterprise-wide AI governance and shadow AI as significant issues. When employees use unapproved AI for work, the company may lack visibility into what information is submitted, where it is stored, whether it is used for training, which subprocessors receive it, what logs exist, and which contractual and security protections apply.
A simple announcement that “unapproved AI is prohibited” may not be sufficient. Where employees have a genuine need for AI-assisted productivity but are not provided with a realistic approved alternative, personal-account use may continue outside the company’s visibility.
A more workable approach is to approve selected enterprise AI services, define the conditions under which they may be used, and prohibit the submission of non-public business information to unapproved AI. Approval should identify not only the service name, but also the relevant enterprise plan, training settings, administrator settings and other technical or contractual conditions on which the approval depends.
8. Connect Information Classification to AI Rules So Employees Know What They May Enter
Rules governing input information are among the most important provisions of any corporate AI framework. A statement such as “confidential information must not be entered into AI” is too abstract if employees cannot determine what the company means by confidential information in a particular workflow.
The AI policy should therefore connect to the company’s existing information-classification framework. Public information, ordinary internal information, personal information, sensitive personal information, trade secrets, non-public M&A information, source code and third-party information received under an NDA may each require different treatment.
For example, public information may be permitted in approved AI as a matter of routine; ordinary internal information may be limited to enterprise environments in which training use is disabled; personal data or highly sensitive trade secrets may require individual approval or a controlled environment; and third-party confidential information may be prohibited unless the relevant contract permits the contemplated processing.
SAKURA Law Office has addressed these issues in greater detail in Legal Update No. 2 on generative AI and Japanese data-protection law and Legal Update No. 4 on generative AI, trade secrets and NDAs.
9. Before Information Is Entered, Confirm That the Company Has the Right to Use It in That Manner
Possession of information does not necessarily give a company the right to submit it to an AI service. Customer-provided data, third-party copyrighted works, business-partner confidential information, joint-research data and licensed content may be subject to statutory or contractual restrictions.
The policy may therefore require users to confirm that the company has the necessary rights or authority to submit the information, or to consult Legal or another responsible function where that authority is uncertain.
This point is particularly important because AI service terms often require customers to represent that they possess the necessary rights in the input and may impose indemnification obligations if a third party brings a claim. Internal rules and vendor contracts should not be treated as separate systems; contractual obligations should be translated into operational controls.
10. Classify Use Cases by Risk and Reserve Additional Review for High-Risk Uses
The Japanese AI Act-related guidelines adopt a risk-based approach under which AI risks are identified and assessed, and controls are calibrated to the degree of impact associated with the field and purpose of use. Corporate AI governance should likewise avoid subjecting every use case to the same level of review.
Summarizing public information, improving writing style or generating ideas presents a different risk profile from screening job applicants, evaluating employees, making credit decisions, providing legal or medical conclusions, making safety-related determinations, automatically responding to customers, approving material contracts or sending messages externally without human intervention.
Companies can therefore require prior review of high-risk uses by reference to factors such as purpose, model, data, effect on decision-making, human oversight, vendor terms and security, while broadly permitting lower-risk routine uses. This allows governance resources to be concentrated where mistakes could cause material harm.
11. Distinguish AI Output Used as Reference Material from AI Output Used as a Final Business Deliverable
The required level of verification depends on how output will be used. Internal brainstorming should not be subject to the same controls as a report delivered to a customer, a contract, an advertisement, a press release, a legal conclusion or a recruitment decision.
The policy should assume that AI output may contain inaccuracies, omissions, bias or content that infringes third-party rights. For material uses, the user should verify appropriate source material and, where necessary, obtain review by a specialist function or responsible manager.
The fact that content was produced by AI should not itself be treated as a defense to inadequate quality control. Where the company chooses to provide the output to customers or third parties, the company should apply the level of review appropriate to the final product.
12. “Human in the Loop” Must Mean More Than Someone Looking at the Screen Once
Human involvement is frequently described as a core element of AI governance, but a policy stating only that “a human will make the final check” may have little practical value.
The company should specify, depending on the use case, who must review what, against which source material, and with what authority to reject or modify the AI output. In high-risk settings, the person approving the AI recommendation must have sufficient knowledge, time and organizational authority to conduct a substantive review. A formal approval click without meaningful examination is not effective human oversight.
Companies should also account for automation bias — the tendency to over-rely on a system’s recommendation. Training, interface design and operating procedures should preserve the user’s ability and responsibility to challenge, reject or correct AI output.
13. Do Not Allow AI Alone to Make Employment Decisions with Material Effects on Employees or Applicants
Using AI to screen candidates, evaluate employees, determine placement or promotion, or support disciplinary decisions raises issues involving personal information, bias and discrimination, explainability and employment management.
As a general principle, companies should avoid allowing a materially adverse employment decision to be completed automatically on the basis of AI output alone. Meaningful human review should remain, and the company should consider whether the data used are relevant to the employment decision, whether the process may reproduce bias, and what records of AI output should be retained.
If violations of the AI policy may result in disciplinary action, the company should also confirm consistency with its work rules and disciplinary provisions. A violation does not automatically justify severe discipline. The exercise of disciplinary authority remains subject to the Labor Contract Act and other principles of Japanese employment law.
14. Address Outputs That May Infringe Copyright, Trademarks, Personality Rights, Reputation or Other Third-Party Interests
AI-generated content may conflict with third-party copyright, trademarks, names, likenesses or other rights. AI may also generate false adverse information about a real person or company, creating potential defamation, reputational or privacy concerns if the content is published or relied upon.
The policy may therefore require appropriate rights clearance for significant outputs intended for external publication or commercial use, prohibit material adverse decisions about real persons or companies based solely on AI responses, and require legal review when questions arise.
SAKURA Law Office has discussed generative AI and copyright in Legal Update No. 3, including the distinction between the training stage and the generation and use of outputs.
15. AI Agents Require Additional Authority Controls Beyond Rules for Chat-Based AI
AI agents may do more than generate text. They may read email, search files, access external services and, depending on configuration, send messages, change data, create reservations, make purchases or perform other actions. A policy designed only to regulate the information entered into a chatbot may therefore be insufficient.
For AI agents, the principle of least privilege is critical. Agents should not receive access rights beyond what is necessary for the defined business function. Actions with significant external impact or limited reversibility — such as payments, contract execution, customer communications, employment actions or deletion of material data — should generally require human authorization.
Because AI agents may read external content, they may also be exposed to prompt-injection and related attacks. Policy language must therefore be combined with technical access controls, sandboxing, logging, execution constraints and other security measures.
16. Align AI Procurement and Contract Review with the Internal Policy
An internal policy stating that only AI services that do not use input for training may be used will fail if the actual contract or administrator settings permit such use. Conversely, even a strong negotiated data-protection clause will not prevent misuse if employees are not told what the contract permits and prohibits.
The approval process for an AI service should therefore review the terms of service, DPA, security conditions, training use, retention, subprocessors, cross-border processing, logging and deletion, and translate those findings into user-facing operating conditions.
Legal Update No. 5 of this series addresses AI service agreements in greater detail, including input, output, intellectual property, liability caps, security, PoCs, APIs and custom development. Internal AI rules and vendor contracts should be designed as components of one governance system.
17. Design Logging Around Purpose and Necessity Rather Than a Default Rule to “Keep Everything”
Logs can be valuable for audits, incident investigations and quality control. At the same time, prompts and outputs may themselves contain personal information, trade secrets or other confidential material, meaning that broad retention can create a new information-security risk.
The company should decide whose activity will be logged, what will be retained, for how long, and who may access it. If employee AI use is monitored, the company should consider the legitimacy and necessity of the monitoring, avoid excessive surveillance and provide appropriate internal notice.
For AI agents, the ability to reconstruct what actions were taken and which human approvals were given is particularly important for incident response and accountability.
18. Incident Reporting Should Be Designed to Stop Harm, Not Merely to Identify a Rule Breaker
If an employee mistakenly submits customer information, a trade secret or third-party confidential information to an external AI service, the speed of the initial response matters. Depending on the service, immediate action may include requesting deletion, removing chat history, disabling an account, revoking access, preserving logs and notifying appropriate internal functions.
The policy should therefore require a user who becomes aware of an incident or suspected incident to report it promptly through a defined channel. That channel should be capable of coordinating Legal, Information Security, the privacy function and other relevant teams.
A culture that disproportionately punishes the act of reporting can create incentives to conceal incidents or delay escalation. Deliberate or serious misconduct may require appropriate consequences, but the first purpose of incident reporting should be containment and prevention of recurrence, not blame allocation.
19. Prepare a Separate Incident-Response Procedure; Do Not Rely on the Policy Alone
A policy clause stating that incidents must be reported is not an incident-response plan. Companies should separately prepare procedures for scenarios such as accidental submission of protected information, compromised AI accounts, erroneous action by an AI agent, misdirected external communications or publication of infringing content.
Where an incident may constitute leakage of personal data, the company must assess whether reporting to the Personal Information Protection Commission and notification to affected individuals are required under the APPI. Where third-party confidential information is involved, contractual notification and mitigation obligations under the NDA may apply. Security incidents may require immediate evidence preservation and technical containment.
Roles should be established in advance: who confirms the facts, who performs legal analysis, who contacts the service provider, and who decides whether and how to report to senior management, customers, regulators or other affected parties.
20. Treat AI Literacy Training as an Internal Control That Makes the Policy Operable
The AI Guidelines for Business emphasize education and AI literacy across actors including AI users. Distributing a policy does not mean employees understand hallucinations, confidentiality risks, copyright, bias, security limitations or the boundaries of an approved AI environment.
It is often useful to distinguish baseline training for all employees from specialized training for higher-risk functions. General users may be trained on input restrictions and output verification, while HR, Legal, engineering, security and other specialist teams receive additional instruction tailored to their specific use cases.
Education should not consist solely of prohibitions. Explaining what employees are permitted to do with approved AI, and how AI can be used safely to improve work, can help reduce shadow AI while preserving productivity gains.
21. Respond to Policy Violations in Proportion to Intent, Impact and Risk of Recurrence
It is reasonable for an AI policy to state that violations may result in consequences, but different violations should not automatically be treated the same. A minor procedural error differs materially from knowingly submitting a critical trade secret to an external AI system.
The company should consider intent or negligence, the sensitivity of the information, likelihood of external disclosure, actual or potential harm, speed of self-reporting and risk of recurrence. Appropriate measures may include additional training, changes in access rights, suspension of AI access or employment-related action where justified.
Any disciplinary action must have an appropriate basis in the company’s work rules or other governing employment documentation and must satisfy applicable principles of Japanese employment law. A statement that “violation of this AI policy may result in discipline” does not make every sanction automatically lawful.
22. Build an Agile Review Process Triggered by Service Changes, Legal Developments and Incidents
The AI Act-related guidelines recognize the speed of technological change and evolving risks and encourage agile improvement of AI governance through iterative review. An internal AI policy should not be frozen based on the services and technology available on the date of adoption.
The company may identify review triggers such as deployment of a new model, introduction of AI agents, material changes to service terms, legal amendments, new regulatory guidance, a significant incident or findings from an audit.
In addition to a periodic annual review, material changes may justify an interim review. Where frequent formal amendment of the core policy is impractical, operating standards and the approved-services register can be updated more dynamically.
23. Do Not Delay Adoption by Trying to Make the First Version Perfect
AI technology and actual workplace use evolve too quickly to predict every future use case in the first policy. The more practical objective is to establish clear minimum boundaries, begin operation, learn from real usage and improve the framework over time.
An initial version may prioritize approved services, prohibited input categories, review of externally used output, prior approval of high-risk use cases and incident reporting. The framework can then expand to AI agents, RAG, internally developed AI, employment use and other more complex areas.
The company should avoid a situation in which drafting becomes a prolonged project while employees continue using AI in the absence of any formal guidance.
24. A Practical Structure for a Corporate Generative AI Use Policy
A practical policy may address purpose and principles, scope, definitions, governance responsibilities, approved AI services, permitted and prohibited information, approval of high-risk use cases, output verification, third-party rights, AI-agent authority controls, logging and records, training, incident reporting, audits, consequences for violations and review of the policy.
The structure should nevertheless reflect the company’s size, industry, AI maturity and existing internal-control framework. Where information-security and confidential-information policies are already sophisticated, it may be efficient to cross-reference them and add only AI-specific controls. In other companies, AI adoption may reveal that information classification or incident response itself requires broader modernization.
The objective should not be to produce one polished document called an “AI Policy.” The objective is to make AI controls work within the company’s overall system of internal governance.
25. Model Clause — Purpose
At the beginning of the policy, it can be useful to state expressly that the purpose is to enable safe and productive AI use rather than merely restrict it. One formulation might be: “This Policy establishes basic rules governing the use of AI in the Company’s business in order to promote the appropriate and effective use of AI while reasonably managing risks including information leakage, infringement of third-party rights, inaccurate information and security incidents.”
Expressly combining responsible use with risk management clarifies that the policy is not merely an instrument of prohibition, but a governance framework intended to enable business value.
26. Model Clause — Approved and Unapproved AI
For approved AI, a company may provide: “Employees may use, for Company business, only AI services separately designated by the Company as Approved AI Services.” Specific service names, plans and conditions can then be maintained in a separate schedule.
For unapproved AI, the company may either prohibit access altogether or adopt a narrower rule such as: “Employees must not submit non-public Company information to any AI service that has not been approved by the Company.” The appropriate boundary should reflect the company’s security posture, business needs and technical controls.
27. Model Clause — Input Information
A core provision might state: “Unless used within an environment and under conditions specifically approved by the Company, users must not submit to AI personal information, trade secrets, material non-public information, information received from a third party subject to confidentiality obligations, or other categories of information designated by the Company.”
A permanent blanket prohibition on every form of personal or confidential information may, however, prevent useful enterprise AI applications. In practice, exceptions based on information category and the security of the AI environment — including individual approval or use within a controlled environment — may be necessary.
28. Model Clause — Verification of Output and Material Decisions
A policy may provide: “Users must not assume that AI output is accurate, complete or lawful and must conduct such verification as is reasonably necessary for the relevant business purpose.”
It may also state: “Recruitment, employee evaluation, disciplinary action, credit decisions, legal determinations and other decisions that may materially affect the rights or interests of an individual or third party must not be made solely on the basis of AI output.” The company should separately determine which functions qualify as high-risk in light of its business.
29. Model Clause — Incident Reporting
An incident clause may provide: “A user who becomes aware of, or reasonably suspects, an accidental submission of protected information, AI output that may infringe third-party rights, an inappropriate external action by AI, account compromise or any other AI-related incident must immediately report the matter to the Company’s designated contact and must not independently delete evidence or conceal relevant facts.”
If users delete logs or history needed for investigation, root-cause analysis may become substantially more difficult. For that reason, reporting and evidence preservation should be addressed together.
30. Frequently Asked Questions
Q1. Is a corporate generative AI policy legally mandatory in Japan?
There is no statute requiring every company to adopt a document with a uniform title and format. Nevertheless, because AI use can implicate personal information, trade secrets, copyright, contracts, cybersecurity and other obligations, companies whose employees use AI for work should generally establish clear internal rules appropriate to their operations.
Q2. Is it sufficient for the policy to cover ChatGPT only?
Not necessarily. AI features are now embedded in office software, meeting platforms, CRM systems, development tools and other business applications. The scope should be considered broadly enough to address embedded AI functions and AI agents where relevant.
Q3. Should all personal information and confidential information be prohibited from AI?
A blanket prohibition is not always the most practical solution. Public AI services and controlled enterprise environments present different risks. The company should distinguish between prohibited use, use subject to specific approval, and use permitted within an approved environment based on the information involved, contractual terms, training use, retention and access controls.
Q4. Must every AI output be reviewed by a human?
The required review should reflect purpose and risk. A low-risk writing-assistance task does not require the same level of verification as an employment decision, a material customer communication or a legal conclusion. A risk-based framework is usually more proportionate.
Q5. Is a ban on employees using personal AI accounts sufficient to eliminate shadow AI?
A prohibition can be part of the solution, but it may not eliminate shadow AI where business demand remains high. Providing an approved enterprise environment and making permitted use clear can, in some organizations, reduce unapproved use more effectively.
Q6. Can an employee be disciplined for violating the AI policy?
Any specific sanction should be assessed in light of the company’s work rules, the nature of the violation, intent or negligence, actual or potential harm and proportionality under Japanese employment law. The existence of a policy violation does not automatically validate every form of discipline.
Q7. Should the policy list specific AI services by name?
The approved services should be identifiable, but names, plans and settings can change frequently. It is often preferable to place the governing principles in the core policy and maintain specific approved services and configurations in a schedule or operating standard.
Q8. Can the same policy govern AI agents?
The core principles can apply, but AI agents require additional controls concerning access rights, external communications, payments, deletion of data and other execution authority. Least privilege and human approval for material actions are particularly important.
Q9. What should the company do after adopting the policy?
The framework must be operated. Employee training, management of approved services, review of high-risk use cases, logging and audits, incident response and monitoring of vendor-term changes should continue. AI governance functions through implementation and review, not through publication of the policy alone.
Q10. Which function should own the policy?
The answer depends on the organization, but Legal should not ordinarily design the framework in isolation. IT, Information Security, Privacy, HR, Internal Audit and key business functions should be involved because AI risk spans legal, technical and operational domains.
31. SAKURA Law Office Perspective on Corporate Generative AI Policies
The most important objective in drafting a corporate AI policy is not to produce elegant policy language. The real objective is to ensure that the written rules are consistent with the AI environments the company actually provides, the vendor contracts it has signed, its information-classification framework, access controls, employee training and incident-response capabilities.
For example, a policy may prohibit employees from entering confidential information into external AI, but if the company provides no approved alternative, non-compliant use may become routine. Conversely, an enterprise AI environment may be deployed without adequate governance if the company assumes that “enterprise” automatically means safe and never examines training use, retention, subprocessors or cross-border processing.
The role of corporate legal practice is not to stop AI use. It is to design the boundaries within which the company can use AI responsibly. Those boundaries cannot be designed by legal doctrine alone; they require an integrated understanding of law, contracts, technology, data, business processes and organizational behavior.
A generative AI policy is therefore not the end state of AI governance. It should be treated as the starting point for a continuing process through which the company deploys AI, identifies problems, improves controls and converts responsible AI use into sustainable business capability.
32. About the Kenshiro Michishita Legal Update Series
SAKURA Law Office maintains a “Legal Update” section on the profile page of Managing Partner Kenshiro Michishita and publishes continuing analysis of legal issues affecting companies and society.
Legal Update No. 1 addressed corporate use of generative AI and AI governance. No. 2 examined generative AI and Japan’s data-protection regime. No. 3 addressed generative AI and copyright. No. 4 examined trade secrets, confidential information and NDAs. No. 5 addressed key terms in AI service agreements.
This sixth installment connects those substantive topics to internal corporate design and explains how AI principles, internal policies, approved-AI frameworks, high-risk use-case review, meaningful human oversight, AI agents, logging, employee training, incident response and policy review can be structured as an operating AI governance system.
Future Legal Updates will continue to address AI agents, AI and employment, AI-enabled new businesses, M&A, international transactions, Web3 and digital assets, crisis management and other issues of practical importance in corporate law.
33. Matters on Which SAKURA Law Office Advises
SAKURA Law Office advises companies on the design and implementation of corporate generative AI policies, AI principles, internal guidelines and broader AI governance frameworks.
Our work may include drafting a new generative AI policy, reviewing existing policies, designing an approved-AI framework, connecting information classification to AI input rules, establishing approval processes for high-risk use cases, designing authority controls for AI agents, aligning internal policy with AI terms of service and data processing agreements, developing employee training, addressing shadow AI, preparing AI incident-response procedures, developing global AI policies covering overseas operations and designing AI governance committees or other accountability structures.
It is not necessary for a company to have complete visibility into its current AI use before seeking advice. We can begin by assessing existing use cases, data, internal policies and deployed services and then design a governance framework appropriate to the company’s business and organizational structure.
34. Contact Us — Corporate Generative AI Policies and AI Governance
For advice regarding corporate generative AI policies, AI principles, internal guidelines, approved-AI programs, AI governance, AI service agreements, AI agents, personal information, trade secrets, copyright, cybersecurity or other AI and technology matters, please contact SAKURA Law Office.
For corporate clients, our support may extend beyond drafting policy language to review of existing information-security rules, privacy rules, work rules, confidential-information policies and AI service agreements, as well as internal approval workflows, employee training, incident response and post-implementation review.
When contacting us, noting that your inquiry concerns a “Corporate Generative AI Policy / AI Governance” will help us direct the matter to the appropriate lawyer promptly.
SAKURA Law Office
Managing Partner: Kenshiro Michishita
4F, Ark Hills South Tower, 1-4-5 Roppongi, Minato-ku, Tokyo 106-0032, Japan
Tel: +81-3-6910-0692
https://sakura-lawyers.jp/en/
Principal Japanese Sources
Agency for Cultural Affairs, materials concerning AI and copyright
Ministry of Economy, Trade and Industry, Trade Secret Management Guidelines and related materials
Labor Contract Act and other applicable Japanese employment laws
Written and supervised by SAKURA Law Office, Managing Partner Kenshiro Michishita
This article provides general legal information as of September 24, 2026, based on Japan’s AI Act-related guidelines, the AI Guidelines for Business Ver1.2, IPA publications concerning generative AI and AI agents, publications of the Personal Information Protection Commission and other relevant authorities, and general corporate legal practice. The appropriate content of a corporate generative AI policy and AI governance framework depends on the company’s business, size, industry, information handled, AI services used, existing internal policies, employment arrangements, security environment and specific use cases. AI-service functionality, terms of use, applicable law and governmental guidance may change. Specific matters should therefore be assessed on the basis of current law, current contractual terms and the actual technical and organizational facts.