# Sample contract clauses for using closed, hosted AI models **Not legal advice.** These are example clause topics and illustrative wording to help a buyer prepare for a conversation with counsel and with an AI model provider. They are not a template to sign as written. Many providers publish standard commercial terms and data processing addenda that already cover some of these points; map each topic to the provider's existing terms first and negotiate only the gaps. Your counsel should adapt every clause to your jurisdiction, your use case and the provider's actual offering. Companion to the Safeguard blog post "How to Build an AIBOM for Closed AI Models, and What to Ask Vendors" (23 October 2026). Defined terms used below (adapt to the agreement): "Provider", "Customer", "Services", "Customer Content" (inputs submitted and outputs returned through the Services), "Model ID" (the identifier Customer sends to select a model), "Documentation". ## 1. Model change notice Provider will give Customer written notice of the deprecation or retirement of any generally available Model ID that Customer has used in the preceding [90] days no less than [6 months] before the Model ID ceases to be available, except where a shorter period is required for safety, security or legal reasons, in which case Provider will give as much notice as reasonably practicable and state the reason category. *Protects:* the lifecycle and retirement date fields in your AIBOM. ## 2. Version stability For each Model ID designated by Provider as a fixed snapshot, Provider will not change the model weights or configuration that respond to requests sent to that Model ID. Any updated model will be made available under a new Model ID. Where a Model ID is an alias that may point to different snapshots over time, Provider will identify it as such in the Documentation. *Protects:* the assumption that your recorded version, evaluations and risk acceptance describe what actually runs. ## 3. Training use of Customer Content Provider will not use Customer Content to train, retrain or improve any model, except where Customer has given explicit written opt in for a specified scope. Any account setting that enables such use will be off by default, and Provider will document which Customer roles can change it. *Protects:* the training use field for each service. ## 4. Retention and deletion Provider will retain Customer Content only for the periods stated in [Schedule X] for each Service and endpoint, and only for the purposes stated there (for example abuse monitoring, legal compliance). Where the parties have agreed a zero data retention arrangement, Provider will identify in writing which endpoints and features it applies to and which may still store data. On termination, Provider will delete Customer Content within [30] days, subject to legal retention obligations disclosed to Customer. *Protects:* the retention fields and the confirmed date in your AIBOM. ## 5. Human review Provider will describe in the Documentation the circumstances in which Customer Content may be accessed by Provider personnel, and the retention period that applies in those circumstances. *Protects:* privacy notice accuracy and the data handling record. ## 6. Sub-processors and location Provider will maintain a current list of sub-processors that process Customer Content, including the country or region of processing, will notify Customer at least [30] days before adding or replacing a sub-processor, and will allow Customer to object on reasonable grounds. *Protects:* the data flow and region fields. ## 7. Model documentation For each generally available Model ID, Provider will make available, and keep reasonably current, documentation describing at least: intended uses and known limitations; the categories of data used for training; the training data or knowledge cut-off date; and a summary of pre-release safety evaluations. Provider will make previous versions of such documentation available on request for Model IDs Customer has used. *Protects:* the provider-asserted layer of your AIBOM and its audit trail. ## 8. Training data information on request On Customer's reasonable written request, Provider will provide, under the confidentiality terms of this agreement, such further information about the data used to train a Model ID as Provider is able to share, or a written statement of the categories of information it cannot share. *Protects:* the record that you asked, and what you were told. ## 9. Regulatory cooperation Provider will provide reasonable information and cooperation that Customer requires to meet Customer's obligations under applicable AI, privacy and consumer protection laws in connection with its use of the Services, including, where Provider is a provider of a general-purpose AI model under Regulation (EU) 2024/1689, the information and documentation Provider makes available to downstream providers, and the location of Provider's public summary of training content where published. *Protects:* your own compliance file. ## 10. Security and model integrity incidents Provider will notify Customer without undue delay, and in any event within [72] hours of confirmation, of any security incident affecting Customer Content, and will notify Customer of any confirmed incident materially affecting the integrity of a Model ID Customer uses. *Protects:* incident response and the event triggers in your AIBOM review cadence. ## 11. Audit evidence Provider will make available to Customer, under confidentiality, its most recent third-party security audit reports covering the Services, and will answer reasonable written security questionnaires no more than [once] per year. *Protects:* the evidence column of your vendor questionnaire. --- Again: this file is not legal advice. Use it to structure discussion with your counsel and the provider.