SAKURA Law Office | Legal Update by Managing Partner Kenshiro Michishita
September 24, 2026
SAKURA Law Office Managing Partner Kenshiro Michishita
Executive Summary
An AI service agreement should not be reviewed as though it were merely an ordinary SaaS agreement with a new feature. The legal and commercial risk depends on what data the customer sends to the service, how the provider and its upstream model or cloud vendors may use that data, what the customer receives in return, whether outputs may be used commercially, how intellectual-property and third-party claims are allocated, what security and audit rights exist, how models may change during the term, and what happens to data and AI assets when the relationship ends. The appropriate contract also depends materially on the transaction structure: use of a general-purpose AI service, API integration, retrieval-augmented generation, fine-tuning, a proof of concept, or custom development. For material AI deployments, legal review should therefore begin with the technical architecture and business workflow, and the contract should be treated as part of the design of the AI governance framework rather than as a document reviewed only at the end of procurement.
SAKURA Law Office is pleased to publish the fifth installment of the Legal Update series by Managing Partner Kenshiro Michishita, entitled “Key Contract Terms Companies Should Review in AI Service Agreements — Risk Allocation and Contracting Practice for Generative AI, APIs, PoCs and Custom Development (2026).”
The way companies adopt AI has changed rapidly. Enterprise use is no longer limited to employees opening a browser and asking questions of a general-purpose chatbot. Companies now integrate foundation models through APIs, connect internal documents through retrieval-augmented generation (RAG), fine-tune models using proprietary data, authorize AI agents to interact with business systems, and conduct proofs of concept before commissioning bespoke AI systems. Each structure creates a different contractual allocation of rights, responsibilities and operational risk.
At the same time, legal teams are still sometimes asked to approve an AI product based only on an order form and a hyperlink to online terms. In practice, the governing terms may be dispersed across a master services agreement, order form, data processing addendum, security schedule, service-level agreement, acceptable use policy, privacy notice, API terms, upstream model-provider terms and even settings in an administrative console. Reviewing a single document may therefore provide an incomplete picture of the legal position.
In February 2025, Japan’s Ministry of Economy, Trade and Industry (METI) published its “Contract Checklist for AI Utilization and Development,” emphasizing that parties should separately analyze the identification, provision, use, external disclosure, rights and processing results associated with both inputs provided to an AI service and outputs received from it. The Checklist also notes that matters such as warranties, indemnification, limitations of liability, termination, governing law and dispute resolution require careful consideration even where they fall outside the checklist’s core matrix.
In March 2026, the AI Guidelines for Business were updated to Version 1.2, and in July 2026 the Information-technology Promotion Agency of Japan (IPA) published practical materials on the safe use of generative AI and AI agents. Contract review in this area can no longer be separated from data governance, cybersecurity, AI governance and the actual business workflow.
This Legal Update summarizes, as of September 24, 2026, the principal contractual issues that companies should consider when procuring or developing AI services, with particular attention to general-purpose AI services, API integrations, PoCs and custom development projects.
1. Start by Determining What the Company Is Actually Buying
The phrase “AI implementation” can describe fundamentally different transactions. A company may simply subscribe to a finished chatbot, integrate a third-party model into its own product through an API, build a RAG environment over internal documents, fine-tune a model using proprietary data, commission a PoC, or engage a vendor to develop and deliver a bespoke AI-enabled system. The relevant contractual issues differ materially across those structures.
METI’s earlier Contract Guidelines on Utilization of AI and Data organized AI transactions principally around development agreements and usage agreements. Those models remain useful, but the market has become more diverse. METI’s materials accompanying the AI Guidelines for Business now expressly recognize general-purpose AI services, customization of third-party AI services and new development as distinct transaction patterns that require careful analysis.
Accordingly, the first step in an AI contract review should not be redlining the agreement. The legal team should first understand the technical architecture and the business process: what leaves the company, whose model processes it, what is returned, where the data and outputs are stored, which systems the AI can access, and who ultimately relies on the result. Only then can the contractual language be assessed against the actual risk.
2. Review the Entire Contract Stack and the Order of Precedence
AI services are frequently governed by a stack of documents rather than a single integrated agreement. An order form may incorporate a Master Services Agreement, online terms, a DPA, a security addendum, an SLA, an Acceptable Use Policy, a privacy policy, API terms and terms imposed by an upstream model provider.
The legal issue is not only which documents are incorporated, but which document prevails if they conflict. A negotiated order form may appear to prohibit training on customer data while online terms retain broad rights to use information for service improvement. If the order of precedence is unclear, the customer may not have secured the protection it intended.
Online terms may also be amended unilaterally during the contract term. For material deployments, key terms concerning data use, confidentiality, security, liability and intellectual property should be fixed in the negotiated agreement where possible, or at least be subject to advance notice, a right to object and, where appropriate, a right to terminate without penalty.
3. Define “Input” Broadly Enough to Capture All Data the Service May Receive
One of the central concepts in an AI agreement is the information provided by the customer to the service. “Input” may include not only text prompts, but uploaded documents, images, audio, source code, documents connected through RAG, data transmitted through APIs, chat histories, embeddings, metadata and information automatically collected by the system.
If defined terms such as “Customer Content,” “Input” or “Customer Data” are drafted too narrowly, some categories of information may fall outside confidentiality obligations, use restrictions or deletion commitments. Logs and automatically generated telemetry are particularly easy to overlook.
METI’s 2025 Contract Checklist begins by identifying the input and then separately examines its provision, use, external disclosure, rights and processing results. The practical exercise for the customer is therefore to inventory the information that may in fact be transmitted and confirm that all material categories are contractually protected.
4. Confirm That the Customer Has the Right to Provide Each Category of Input
Before a company sends information to an AI provider, it must have the legal right to make that information available for the proposed processing.
The analysis may be straightforward for ordinary documents created solely by the company, but it becomes more complex for customer data, materials received under an NDA, licensed content, employee information, third-party copyrighted works, jointly developed research data or information subject to sector-specific restrictions.
Provider agreements often include broad customer representations that the customer holds all rights necessary to submit the inputs and broad indemnities for third-party claims arising from those inputs. If the company lacks internal controls over what employees may submit, a seemingly standard representation can concentrate substantial risk on the customer side.
5. Determine Whether Inputs May Be Used for Model Training, Service Improvement or Other Secondary Purposes
One of the highest-priority issues in an AI service agreement is secondary use of customer inputs.
METI’s Contract Checklist distinguishes among arrangements in which inputs may be used for general model training, used only to provide or improve services for the particular customer, or not used for training at all. The risk profile differs significantly among those structures.
The review should therefore go beyond the binary question “Is our data used for training?” and examine each relevant purpose separately, including foundation-model training, general service improvement, customer-specific fine-tuning, quality assurance, abuse monitoring, security analysis and statistical analytics.
Where the customer may submit personal data, trade secrets or sensitive technical information, the agreement should be reviewed carefully to ensure that the provider cannot use those materials to improve services for unrelated customers unless that use is specifically intended and lawful. In material cases, reliance on a user-controlled opt-out may be insufficient; the restriction should be stated as a contractual obligation.
6. Draft Confidentiality Provisions Around the Actual AI Data Flow
A generic confidentiality clause may not be enough to protect sensitive information in an AI environment. A foundation-model provider, cloud provider, subprocessors and support personnel may all participate in the processing chain.
The agreement should confirm that relevant inputs fall within the definition of confidential information, that confidential information may be used only for the contracted service, that access is limited on a need-to-know basis, that equivalent obligations are imposed on subprocessors, and that information is returned or deleted when the purpose or contract ends.
Where the service may process trade secrets or information received under third-party NDAs, the agreement should address concrete processing activities rather than relying solely on an abstract promise to “protect Confidential Information.” Training use, log retention, human review, subprocessors and deletion all require separate attention.
7. Where Personal Data Is Involved, Read the AI Terms and the DPA as One Contractual Framework
If the AI service processes personal information, the customer should review the online terms together with the DPA and other data-processing provisions rather than treating them as separate exercises.
Japan’s Personal Information Protection Commission has cautioned that where a business inputs personal data into a generative AI service without the individual’s consent and the data is handled for purposes beyond generating the requested response, the processing may create issues under the Act on the Protection of Personal Information. The business should therefore confirm, among other matters, whether the data is used for machine learning.
Contract review should address the purpose of processing, the legal characterization of outsourcing or third-party provision, cross-border processing, subprocessors, data location, security controls, incident notification and deletion at the end of the relationship.
The fact that a service is cloud-based does not, by itself, determine whether there is a “provision” or “outsourcing” of personal data under Japanese law. The contractual description should match the provider’s actual ability to access and handle the data.
8. Define the Output and the Scope of Permitted Use
The meaning of “output” varies widely across AI transactions. It may consist of a chat response, generated image, generated code, analytical result, embedding vector, trained model, evaluation report or fine-tuned model.
METI’s Contract Checklist applies a structure similar to that used for inputs, separately considering the identification, provision, use, external disclosure, rights and processing results associated with outputs.
If the company plans to incorporate outputs into advertising, products, customer deliverables or proprietary software, the contract should be reviewed for commercial-use rights, modification, reproduction, distribution, disclosure and sublicensing. A PoC license that permits internal evaluation may not permit production use or onward sale.
9. Ownership of Output Is Not the Same as an Unrestricted Right to Use It
AI agreements sometimes state that outputs are “owned” by the customer. That statement does not resolve every legal issue.
Whether copyright or another intellectual-property right subsists in an AI-generated work depends on the facts, including the degree of human creative contribution. A contractual assignment from the provider also does not eliminate the possibility that an output infringes a third party’s copyright, trademark, publicity right or other interest.
The review should therefore consider not only ownership, but also whether the provider may reuse outputs, whether identical or similar outputs may be generated for other users, whether any exclusivity is granted and whether the service imposes commercial-use restrictions. Where AI-generated material is intended to become a core brand or product asset, this analysis is more important than a simple ownership clause.
10. Examine Warranties and Indemnities for Third-Party Intellectual-Property Claims
If generated text, images or code infringe a third party’s intellectual-property rights, the allocation of that risk can become one of the most important provisions in the agreement.
Some providers offer indemnification for specified intellectual-property claims, but the protection may cover only the service itself and not individual outputs. Coverage may also depend on using a specified model, enterprise plan, safety setting or other provider-defined condition.
Conversely, the customer may be required to indemnify the provider for claims arising from customer data or prompts. It is therefore not enough to confirm that the contract contains an “IP indemnity.” Counsel should identify whose conduct is covered, which rights are included, what procedural conditions apply, what exclusions exist and whether the indemnity is subject to a monetary cap.
11. Review Disclaimers Concerning Accuracy, Hallucination and Fitness for Purpose
Generative AI produces probabilistic outputs and cannot eliminate all errors. Many standard AI agreements disclaim warranties of accuracy, completeness, non-infringement or fitness for a particular purpose.
A broad disclaimer may be acceptable for low-risk drafting assistance, but it may be inadequate where AI is embedded in screening, diagnosis, pricing, legal, financial, HR or safety-critical functions.
For material use cases, the parties may need to agree on evaluation datasets, performance metrics, repeatability, response times, error rates, review procedures and human oversight. In custom development, a promise that the model will always achieve a fixed level of accuracy may be technically unrealistic; the contract should instead define the evaluation method, test conditions, remediation process and acceptance criteria with precision.
12. Consider Model Changes, Deprecation and Service Continuity — Not Only the SLA
Traditional SaaS agreements focus heavily on uptime and response-time SLAs. AI services add another layer of operational risk because a foundation-model change can alter answer quality, output behavior, context length and functionality even where the service remains “available.”
If the customer has built a business process around the performance characteristics of a particular model, unilateral deprecation or replacement may materially affect operations.
For critical systems, the agreement should address advance notice of model retirement, migration to replacement models, API changes, rate limits, pricing changes, backward compatibility and remedies for material performance degradation.
Where the customer incorporates the AI service into a service offered to its own customers, it should also compare the SLA it promises downstream with the SLA and remedies it actually receives from the upstream AI provider.
13. Allocate Rights in Fine-Tuned Models, Additional Training Results and RAG Assets
Where the customer optimizes AI using its own data, a simple input/output framework may be insufficient. Intermediate assets may include a fine-tuned model, training dataset, embeddings, vector database, prompt templates, evaluation data and tuning know-how.
METI’s earlier AI contract guidance recognized that where additional training produces a reusable model, the parties should address how data contributions, sensitivity, cost, know-how, permitted use and responsibility affect the allocation of rights and reuse.
The agreement should state who controls each asset, whether the provider may reuse it for other customers and what the customer may export or retain after termination. A customer that contributes strategically important data but receives no continuing rights in the resulting model or related assets may become heavily dependent on the vendor.
14. Understand Non-Exclusivity and the Possibility of Similar Outputs for Other Customers
Generative AI may produce similar results in response to similar inputs. Receiving an output does not ordinarily guarantee that another user will never receive a comparable result.
That may be immaterial for routine drafting or summarization, but it matters if the company intends to treat AI-generated product names, advertising concepts, algorithms or strategy proposals as proprietary assets.
In custom-development projects, the contract should also distinguish between the vendor’s right to reuse general know-how and the customer-specific data, models, prompts, workflows and business logic that must not be reused for other clients.
15. Map the Supply Chain: Subprocessors, Foundation-Model Providers and Cloud Providers
The company’s contractual counterparty may not have developed the underlying AI model. The service may combine a third-party foundation model, cloud infrastructure and other data-processing services.
This creates a multi-layered structure for data handling, processing locations, usage rights and security responsibility.
The customer should review the categories of material subprocessors, notice of additions or changes, objection rights, flow-down obligations and the provider’s responsibility for its subcontractors.
It should also understand whether changes to the upstream model provider’s terms could affect the customer’s own contractual position or the continuity of the service. Material upstream dependencies should be identified before the AI becomes operationally critical.
16. Security Clauses Should Address AI-Specific Risks and Logging
An AI-service security clause should address not only conventional cybersecurity controls, but also the way AI is actually used.
METI’s Contract Checklist identifies system security levels, audit provisions and log retention as concrete areas requiring attention. IPA’s 2026 materials on generative AI and AI agents likewise emphasize enterprise AI governance, shadow AI, lifecycle management and AI-specific risk assessment.
The agreement should address encryption, access controls, authentication, vulnerability management, tenant separation, incident notification, recovery, log retention and forensic cooperation. Where AI agents can act on external systems, the company should also consider execution permissions, prompt-injection risk, unauthorized disclosure and the allocation of responsibility for autonomous actions.
17. Design Audit Rights for Practical Effectiveness, Not Merely for Formal “Inspection” Rights
For major cloud and AI providers, unrestricted on-site audits by every customer may be unrealistic. Audit provisions should therefore be assessed by their practical ability to provide assurance rather than by whether they use the word “inspection.”
Third-party certifications, independent audit reports, security questionnaires, disclosure of findings, enhanced information rights following a serious incident and cooperation with regulators may be combined to create a workable assurance framework.
For highly sensitive data or regulated industries, standard certifications may not be sufficient. The customer should determine whether its own statutory or supervisory duties can be discharged using the information it is contractually entitled to obtain.
18. Unilateral Amendment Clauses Require Particular Attention in AI Services
AI products change quickly. Online terms, model specifications, pricing, features and data-use conditions may all change during the contract term.
METI’s Contract Checklist treats amendment of terms as an independent review item. If a company embeds AI in a material business process, a provider’s ability to change core conditions unilaterally can undermine the risk assessment on which internal approval was based.
For significant changes, the customer should consider requiring advance notice, restricting unilateral adverse changes to core terms on data use, confidentiality, security and liability, and preserving a right to terminate without penalty where a change is unacceptable.
19. Review Limitations of Liability for Carve-Outs and Reciprocity — Not Only the Cap
AI service agreements often cap provider liability at fees paid during a defined period and broadly exclude lost profits, indirect damages and special damages.
Where the subscription fee is modest but the service processes highly sensitive data or supports a critical business function, the potential loss may be radically disproportionate to the standard cap.
Negotiation should therefore consider whether confidentiality breaches, personal-data incidents, IP indemnities, willful misconduct or gross negligence should be subject to a separate cap or excluded from the general limitation. The enforceability and appropriate level depend on governing law and the transaction, but counsel should at minimum check for asymmetry — for example, unlimited customer indemnities alongside a very narrow provider liability cap.
20. Build Exit, Portability and Deletion Into the Contract from the Beginning
Over time, an AI service may accumulate prompts, chat histories, evaluation datasets, RAG databases, fine-tuned models, configuration data and logs. If those assets cannot be retrieved at termination, migration to another provider may become difficult or commercially impracticable.
The agreement should address export rights, data formats, transition assistance, retention periods, deletion deadlines, deletion from backups and, where appropriate, deletion certification.
In custom development, the parties should also define delivery of source code, models, configurations, documentation, training data and evaluation data. Exit planning at the beginning of the relationship is one of the most effective ways to reduce vendor lock-in and business-continuity risk.
21. Review Governing Law, Dispute Resolution and the Availability of Injunctive Relief
Standard terms of overseas AI vendors often select foreign law and foreign courts or arbitration. That may be acceptable for low-value, low-risk use, but a company that entrusts critical data or business functions to the service should consider the cost and effectiveness of enforcing its rights abroad.
Confidential-information misuse and intellectual-property infringement may require immediate injunctive or conservatory relief rather than damages after the fact. The dispute-resolution clause should not inadvertently prevent a party from seeking urgent interim measures where necessary.
Negotiation priority should not be determined solely by subscription price. A low-cost AI service may create high contractual risk if the data or business process involved is strategically important.
22. A PoC Agreement Should Not Prematurely Impose Full Production Obligations — But It Must Still Protect Data and Results
AI projects often benefit from separating assessment, PoC, development and additional training because specifications and achievable performance may be uncertain at the outset. METI’s earlier AI contract guidance promoted an exploratory, phased development approach for this reason.
The purpose of a PoC is ordinarily to test whether a proposed combination of data and technology is likely to produce useful results. Imposing the same completion obligation or accuracy guarantee that would apply to a finished production system may be inconsistent with the nature of the exercise.
However, a PoC agreement should not be reduced to a bare purchase order. Data-use rights, confidentiality, rights in PoC results, reuse of models and know-how, conditions for moving into production, fees, and deletion or return of data if the project ends should all be addressed from the PoC stage.
23. Custom Development Requires Contractual Rules for Acceptance, Performance Evaluation, Change Control and Ongoing Operations
A bespoke AI system or AI-enabled application cannot usually be governed by standard SaaS terms alone.
AI performance may depend on training data, model behavior, prompts and external services, and the system may not reproduce exactly the same output in the way traditional deterministic software does. Acceptance should therefore not be reduced to a single metric such as “95% accuracy.” The contract should define evaluation data, measurement methodology, acceptable error, re-testing, remediation procedures and out-of-scope conditions.
After launch, the parties should also allocate responsibility and cost for model updates, data drift, changes to upstream foundation models, incidents, retraining and ongoing monitoring.
A project in which the parties negotiate ownership during development but fail to define operational responsibility may simply move the dispute from the build phase to the production phase.
24. Prioritize Contract Negotiation by Business Impact, Not Merely by AI Spend
It is not realistic to negotiate bespoke terms for every AI tool used by a large organization. The depth of legal review should therefore be proportionate to the risk of the use case.
A low-cost service used only to polish public-facing text presents a different risk from a service that processes large volumes of customer data or powers a core customer product.
Useful prioritization factors include the sensitivity of the information, whether personal data is involved, the significance of decisions influenced by AI, impact on external customers, degree of vendor dependency, substitutability and the likely scale of loss if the service fails or data is misused.
For material deployments, legal, information-security and privacy teams should become involved during product selection, not only after the business team has chosen the vendor and is ready to sign.
25. Frequently Asked Questions
Q1. If an AI service is marketed as “enterprise,” can we safely accept the standard terms?
Not necessarily. Enterprise branding does not resolve the legal issues. The company should review the intended use, data inputs, training and secondary use, personal data, confidentiality, third-party IP, liability limitations, subprocessors, cross-border processing and unilateral amendment rights. Where material data or business functions are involved, negotiated terms or an enterprise addendum may be appropriate.
Q2. Is it enough if the provider says that customer data is not used for training?
No. Training is only one part of the data lifecycle. The company should also review retention, provider and subprocessor access, quality-control or security uses, logging, processing locations and deletion after termination. A “no training” commitment is important but does not, by itself, resolve confidentiality, privacy or trade-secret risk.
Q3. If the terms say that the output belongs to us, can we use it freely for commercial purposes?
Not necessarily. Ownership is distinct from third-party infringement risk, commercial-use restrictions, acceptable-use rules and the possibility that similar outputs may be generated for others. Companies should review both the contractual license and the underlying IP risk before using AI-generated material as a significant business asset.
Q4. If an AI error causes loss to our customer, can we recover the full amount from the AI provider?
That depends on the contract. AI providers often disclaim accuracy and cap liability by reference to fees paid. A company’s liability to its own customer may therefore exceed what it can recover upstream. For important customer-facing use cases, the company should assess this liability gap before deployment.
Q5. Can a PoC be handled with a short purchase order only?
Sometimes, but even a PoC should ordinarily address data use, confidentiality, rights in results and models, vendor reuse, fees, conditions for production deployment, and return or deletion of data if the project ends. A failed PoC should not leave the parties uncertain about who may retain or reuse the data and intermediate assets.
Q6. What should we do if the provider can amend its online terms at any time?
For material use cases, contractual change management is important. The company should seek advance notice of significant changes and, where appropriate, a right to object or terminate. Internally, legal or system owners should also monitor material changes to terms that could affect the approved risk profile.
Q7. What deserves particular attention when contracting with an overseas AI provider?
The company should review processing locations, cross-border personal-data issues, subprocessors, intellectual-property terms, limitations of liability, governing law, courts or arbitration, emergency relief and unilateral changes. If sensitive technical information is transmitted overseas, security-export-control issues under Japan’s Foreign Exchange and Foreign Trade Act may also need separate analysis.
Q8. Which provisions should be reviewed first in an AI service agreement?
There is no universal ranking. A practical starting point is the customer’s data and intended use: input-use rights, confidentiality and personal data, output-use rights, third-party IP, liability limitations, security, amendment rights and exit. For material deployments, the review should extend to PoC or development results, performance evaluation, SLA, subprocessors and portability.
26. The SAKURA Law Office Perspective on AI Service Agreements
The most dangerous mistake in AI contract review is to treat the service as merely “a slightly unusual SaaS product.” In an AI transaction, the customer’s data may be used to train or improve models and may contribute indirectly to services delivered to others. At the same time, the customer’s outputs may carry uncertainty as to accuracy, exclusivity and intellectual-property protection. The agreement must therefore trace the flow of both data and results more carefully than many conventional SaaS contracts require.
There is also a limit to what legal can accomplish by redlining terms in isolation. A contract may prohibit training on customer data, yet the administrative console may still enable a setting inconsistent with that promise. A contract may require logging that the product does not technically support. A vendor may promise deletion that cannot be implemented in the expected manner. The legal position and the technical facts must match.
The proper role of corporate legal practice is not to catalogue every possible risk and thereby prevent AI adoption. It is to identify the risks that matter for the company’s particular use case and manage them through a combination of contract, technical configuration, internal policy, access controls and human review.
An AI agreement should therefore be treated not as paperwork to be signed at the end of procurement, but as one component of the architecture through which AI is safely integrated into the business.
27. About the Kenshiro Michishita Legal Update Series
SAKURA Law Office continues to publish the Legal Update series on the profile page of Managing Partner Kenshiro Michishita, addressing legal issues of practical importance to companies and society.
The first installment addressed corporate use of generative AI and AI governance in Japan; the second examined generative AI and Japan’s Act on the Protection of Personal Information; the third addressed generative AI and copyright; and the fourth examined generative AI, trade secrets, confidential information and NDAs.
This fifth installment connects those data, confidentiality and intellectual-property issues to the contracts companies actually sign when procuring or developing AI services, and addresses inputs, outputs, intellectual property, security, liability, PoCs and custom development as an integrated contractual framework.
Future Legal Updates will continue to address internal AI policies, AI agents, AI and employment, AI-enabled new businesses, M&A, cross-border transactions, Web3 and digital assets, crisis management and other areas of corporate law in which technology and legal responsibility increasingly intersect.
28. How SAKURA Law Office Can Assist with AI Service Agreements
SAKURA Law Office advises companies on contracts arising from the adoption, use and development of generative AI and other AI services.
Our work may include reviewing AI terms of service, Master Services Agreements, order forms, DPAs, SLAs and security schedules; negotiating AI-provider agreements; revising provisions on model training and confidentiality; analyzing commercial use and intellectual-property issues relating to AI-generated outputs; negotiating indemnities and liability limitations for third-party claims; structuring API, RAG and fine-tuning arrangements; drafting PoC agreements, AI development agreements and maintenance agreements; and addressing personal-data, trade-secret and cross-border issues arising from AI implementation.
Companies may consult us before selecting a vendor or finalizing the technical architecture. By reviewing the proposed use, data, technical structure, business criticality and vendor terms together, we can identify which issues warrant negotiation and which risks can instead be managed through configuration, governance or operational controls.
29. AI Service Agreement Inquiries
For advice concerning generative-AI terms of service, API agreements, PoCs, custom development, AI service agreements, intellectual property, personal data, trade secrets, cybersecurity and contractual allocation of AI-related risk, please contact SAKURA Law Office.
For corporate clients, our support includes not only contract review and negotiation, but also pre-deployment risk assessment, comparison of contractual conditions among vendors, alignment with internal AI policies, approval workflows, post-deployment contract management and response to changes in online terms.
When contacting us, noting that your inquiry concerns an “AI Service Agreement” 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
Ministry of Internal Affairs and Communications / METI, “AI Guidelines for Business Ver1.2”
METI, materials concerning the “Contract Guidelines on Utilization of AI and Data”
Written and supervised by SAKURA Law Office, Managing Partner Kenshiro Michishita
This article provides general legal information as of September 24, 2026, based on METI’s Contract Checklist for AI Utilization and Development, AI Guidelines for Business Ver1.2, the Contract Guidelines on Utilization of AI and Data and related materials, publications of Japan’s Personal Information Protection Commission and IPA, and general contracting practice. Appropriate contractual terms and allocation of risk in any AI transaction depend on the technical architecture, intended use, data involved, transaction structure, governing law, bargaining position of the parties, regulatory context and business criticality. AI terms of service, model specifications, pricing, data-use conditions and security conditions may change over time. Specific matters should therefore be assessed on the basis of current law, current contractual terms and the actual technical facts.