AI Vendor Due Diligence and Contract Review: Points to Check for Adopting Companies and Providers
Hello, I'm Noriaki Asato, Representative Attorney at LegalAgent.
A business unit submits an approval request saying, "We want to introduce this AI tool to improve operational efficiency." The attached materials are a price list and a brochure describing the features, and the security checklist is the same form used for previous SaaS introductions. As a result, the legal team is unsure which items to check and how deeply. It is not unusual for consultations about reviewing the introduction of AI services to begin with everyday situations like this.
Even in a typical SaaS review, we check whether data is sent externally and stored in the cloud. However, tools such as generative AI, meeting-minutes AI, and contract review AI are often connected directly to a company's confidential information and to data at the core of its decision-making. Checking only formal conditions such as usage fees, contract term, and service levels set in the SLA will not catch the issues specific to AI services. A review is sufficient only once it goes further and confirms whether input data will be used to train the model, how much will be recorded in logs, and who bears responsibility for errors in the generated results.
In addition, public guidelines on AI, related laws such as the Act on the Protection of Personal Information, and the terms of the services being used may be revised, so it is essential to check them against the latest published information each time a tool is introduced.
Preconditions to Settle Internally Before the Review
Before sending a questionnaire to the vendor or reading the terms in detail, you first need to clarify your own company's conditions of use. If the internal premises are not settled, there is no yardstick for evaluating the vendor's answers, and the process ends with simply receiving the materials submitted.
The following five items should be confirmed in advance.
- Whether the data to be input includes confidential information, personal information, or copyrighted works of third parties
- Who will use the generated output, which work it will be applied to, and the scope within which its use is permitted
- Whether you can explain, through the contract and internal rules, whether there is external transmission, use for training, log retention, or reuse
- Whether the premises of the service's terms of use, the privacy policy, your internal rules, and individual contracts are consistent
- Whether a person responsible for the final decision and a checking procedure are in place, so that AI answers are not adopted as is
Points of Focus in AI Vendor Due Diligence
AI vendor due diligence refers to the process of checking the service provider's technical level, data management system, and proposed contract terms before introducing or selling an AI service. Unlike a review of general software, for AI services the scope of checks extends to the characteristics of the model, the use of data for training, and the rights in generated output. For example, even among services all described as "meeting-minutes AI," the risks to be borne differ greatly depending on whether the service processes audio data temporarily and discards it, retains it to improve the service, sends data to an external large language model, processes it on overseas servers, or allows the vendor's administrators to view logs directly. For contract review AI as well, the assessment changes depending on whether contract text is excluded from model training, reused as statistical data, or kept in separate knowledge bases for each customer. In addition to product labels such as "minutes tool" or "review tool," check the specific flow of data (data flow) individually.
This checking work cannot be completed by the legal team alone. Confirming the technical data flow requires the expertise of the IT department and security staff, and evaluating business necessity and impact requires the perspective of the business unit. When the tool will be used in high-risk areas such as finance, healthcare, or HR and labor, it is realistic to include the department responsible for personal information protection and senior management and to check at a depth proportionate to the importance of the work.
Allocation of Responsibility and Data Flow in Contract Review
Reviewing AI vendor contracts tends to be problematic because AI is not just a place to store data; it becomes involved in the company's decision-making process itself. If AI drafts an incorrect response to a customer and damage results, how far is the vendor exempt from liability? If an employee mistakenly inputs confidential data, does the vendor have a procedure for deleting it? If a model change without prior notice significantly reduces the accuracy of the output, does the user company have a way to terminate the contract? If the contract does not clearly set out how these situations will be handled, the burden of negotiating a response after an incident, while checking the governing law and what was agreed, increases. Also check how the SLA applies when the service is down.
Note too that multiple businesses stand behind an AI tool. Where the application provider, the developer of the foundation model, and the cloud provider handling hosting work together, reading the body of the contract alone will not tell you where the data ultimately flows. In addition to the contract clauses, compare the security white paper, the settings in the admin console, and the list of subcontractors (sub-processors) to check that what was agreed and the actual data processing do not contradict each other.
Detailed Review Items Aligned with the Work
When carrying out the review concretely, check items following the flow of the work. First is the nature of the input data and whether it is used for training. If confidential information or data entrusted by customers will be input, check that this does not conflict with subcontracting clauses or restrictions on use for other purposes in contracts with those customers. Requirements regarding the provision of personal data to third parties or to foreign countries should also be checked separately according to how the data is handled. Whether data is used for training is an especially important item. Even if a salesperson says "data will not be used for training," check whether this matches the wording of the contract and the default settings in the admin console. If there is a gap between what was said during sales and what the contract clauses state, do not simply assume that one automatically prevails; check the order of precedence under the contract and resolve the discrepancy by stating important agreed matters expressly in the contract. Labels such as "anonymization is performed" or "this is a corporate plan" do not by themselves resolve all legal issues.
Next are data retention periods and deletion procedures. Check where and for how long input data, output, prompt logs, and backup data are stored, and whether they can be deleted immediately through an administrator's action. Also note the number of days required to return or destroy data upon termination of the contract.
On security, check authentication features, access rights management, and the country where the data center is located, and also bring into view resistance to prompt injection, prevention of access to data beyond one's authorization, control of unintended external transmission, and measures to prevent abuse of authority by the vendor's administrators.
For the handling of output, check the scope of clauses disclaiming any guarantee of accuracy, the checking obligations imposed on users, indemnification clauses for third-party infringement claims, prohibited uses, the treatment of rights in the output, and whether commercial use is permitted. On the premise that AI output may contain errors, decide at which stage of the work procedure to place human checks, in parallel with the contract review.
It is also important to check the prior notice period for model changes, revisions to the terms of use, and fee changes, as well as termination rights and how data can be migrated to another service. Even at the PoC (proof of concept) or trial stage before formal introduction, data use conditions may differ from the main contract, so agree in advance on restrictions on the data that may be input, whether anonymization is required, deletion of data after the trial ends, and ownership of rights in deliverables from the trial. Putting not only price and convenience but also whether data is used for training, the level of security, and the conditions for notification of incidents into the same comparison table for internal approval helps prevent gaps in understanding between departments.
Differences Between the Positions of Adopting Companies and Providers
For a company introducing an AI tool, the main concerns are protecting input data and avoiding excessive operational dependence on a particular service. For work where errors in output could damage the company's external credibility, it is necessary to prepare a human verification system and alternative means in case of failure. You must also not forget to consider whether the use would breach confidentiality obligations to your own customers or restrictions on transfers abroad.
On the other hand, for a company providing an AI service to customers, the central task is to present the limitation that the completeness of the output cannot be guaranteed and to clarify the scope of the user's responsibility in the contract. Selling while overstating features or safety can lead to claims of breach of contract or tort liability when problems arise. Also check whether there is a gap between the terms of use of the foundation model and cloud provider you have adopted and the terms you promise your customers. Where your promises on training use, infringement indemnification, or liability caps are broader than the terms of the foundation model and the like, also decide whether your company can bear that difference.
What both sides should keep in mind is not to treat a contract for an AI service as finished once it is signed. Because this is a field in which terms and models change frequently, it is desirable to maintain a practice of reconfirming the data flow and conditions of use at periodic contract renewals and when the number of using departments expands.
Comparing the Contract with How the Tool Is Used
AI can be used to read terms of use, security materials, and sub-processor lists and pick out items to check regarding input data, training use, and transfers abroad. However, handing over only the contract will not produce a judgment that fits the department using the tool, the information input, and how widely the output will be made public. A person checks the current terms and admin console, and decides, following the flow of the work, who approves use, who checks the output, and who explains in the event of an incident.
Reviewing the introduction of an AI service is not limited to a one-off consideration of an individual tool; it is an opportunity to review the company's internal workflows and its contractual relationships with business partners as a whole. LegalAgent provides support in the following areas of expertise, including building ongoing business relationships and contract investigations in M&A.
Services connected to this topic