AI Agreements for Businesses in Israel: Data, IP Rights, Liability and Governance

A company implementing an artificial intelligence system in its customer service operations may give the provider access to correspondence, commercial documents and personal information. If the system is also authorized to respond to customers, modify orders or operate other services, it may take actions with financial and legal consequences. Terms of use permitting the provider to operate the model do not necessarily answer who may use the information, what happens when the model changes or who bears the loss caused by an incorrect action.

AI agreements raise issues of computer and internet law, commercial contracts, privacy and intellectual property. The arrangements required depend on the activity: purchasing an existing AI tool, developing a bespoke system, integrating a model through an API, deploying an autonomous agent or allowing employees to use external tools. In each case, it is necessary to determine who controls the information and the system’s actions, who gives the contractual undertakings and what remains with the business when the relationship ends.

A Dedicated AI Agreement, a SaaS Addendum or an Acceptable Use Policy?

An agreement with an AI provider need not be a stand-alone document. Where a company purchases access to an existing system, the relationship may be governed by a master services agreement, an order form and addenda dealing with data processing, confidentiality and the use of AI. Standard SaaS and cloud-service arrangements remain relevant to payment, service availability, support and termination. An AI system, however, raises questions not necessarily addressed by a commitment merely to provide access to software: whether inputs may be used to improve the model, whether results are reproducible and whether the provider may replace a model in a way that changes service quality.

Where the provider is engaged to develop a system, train a model or create a component specifically for the customer, the commissioned work, training data, acceptance criteria and rights in the developed components must be separately defined. These questions overlap with those arising under software development agreements, but they do not end with delivery of the code. A model-based system may change after development is completed as a result of updates, additional training or a change in the underlying model provider.

A different type of arrangement arises where the company does not procure an AI system itself, but allows its employees or contractors to use external tools. In that situation, negotiating the AI provider’s terms may not always be possible. Practical protection requires employment and service agreements, together with internal policies, to specify what information may be entered into the tools, which outputs require review and who may approve a new tool.

Israeli Law and Regulation Applicable to AI Use

Israel does not currently have a single overarching law regulating all business uses of artificial intelligence. Rights and obligations arise, among other sources, under contract law, the Protection of Privacy Law, 5741-1981, the Copyright Law, 5768-2007, the Commercial Torts Law, 5759-1999, and tort and consumer protection law where applicable. Additional sector-specific requirements may apply in regulated industries. The Israeli government’s 2023 AI policy document favored sectoral regulation at that stage and set out principles for responsible use, but it does not itself establish a uniform set of contractual obligations for all private companies.

Amendment No. 13 to the Protection of Privacy Law, which entered into force in August 2025, updated statutory definitions and expanded enforcement mechanisms. A company transferring personal information to an AI provider must assess the purposes of the processing and the parties’ respective roles in light of what they actually do, rather than relying solely on the labels used in the agreement. Draft guidance published by the Israeli Privacy Protection Authority for public comment in April 2025 explains the Authority’s position on the application of privacy law to the development, training and operation of AI systems. That document was a draft released for comments, not new legislation.

Operations involving the European Union require separate consideration of the EU Artificial Intelligence Act and the GDPR where personal data are processed within its scope. The AI Act distinguishes between providers and deployers, and between different categories and risk levels of AI systems. Its territorial application may extend beyond the EU, for example where a provider places a system on the EU market or where the system’s output is used in the EU in circumstances covered by Article 2. The mere fact that an Israeli company uses a product supplied by a European provider does not establish that every provision of the AI Act applies to that company. An international agreement should identify each party’s applicable compliance obligations and responsibility for providing the information and documentation necessary to meet them.

Due Diligence on the AI Provider and Its Own Rights

A supplier marketing an AI solution may have developed only the user interface while relying on another company’s model, an additional hosting provider and external tools for search and automated actions. The customer should identify the parties in this service chain and distinguish the direct supplier’s promises from the terms imposed on it by the underlying model operator. The review should also identify who receives the information, where it is processed and whether subcontractors may retain or use it.

Before contracting, the customer should examine the model provider’s terms, usage restrictions, documentation describing the system’s capabilities and security measures, and the ability to obtain records of product changes. Where a supplier offers commercial use of outputs or promises indemnification against infringement claims, it is important to establish whether the supplier has suitable licences and corresponding commitments from the providers of the technology and components on which it relies. A broad contractual promise from a technology intermediary may leave the customer exposed if there is no adequate protection further up the supply chain.

There is a difference between a service in which the model runs in an environment dedicated to the customer and a public service that transmits data to shared infrastructure. Terms such as “Enterprise”, “Private” and “Dedicated” do not, by themselves, establish whether inputs are used for training, who may inspect them, whether usage logs are retained or which external services can access them. The agreement should set out the actual permissions and restrictions.

Customer Inputs and Model Training

One of the most sensitive contractual issues is the separation between processing information to deliver the service and the supplier’s independent use of that information. To provide the service, the supplier may process documents, user instructions, customer information and intermediate results. This does not automatically authorize the supplier to use those materials to train general-purpose models, test new products or serve other customers. If additional uses are permitted, the agreement should define their purposes, the categories of information concerned and the scope of the permission.

The contractual definition of customer information should cover not only files uploaded to the system, but also prompts, conversations, search results, logs, knowledge bases from which the system retrieves information and data derived from use. In a system combining an external document repository with a language model, such as a retrieval-based architecture, information may be held in several components, including indexes and vector representations. The parties should determine what is retained in each component, who can access it and what is deleted when the service ends.

A general undertaking not to “sell data” is not equivalent to a prohibition on training. Even a commitment not to train the underlying model may fail to cover the use of customer examples to improve ranking mechanisms, perform quality testing or develop new features. The agreement can prohibit such uses without express authorization, introduce approval requirements for new uses, restrict access by human reviewers and require deletion of information retained for operational purposes. If the supplier seeks to use aggregated or anonymized data, it is necessary to define what makes data non-identifiable and how combinations of data that could re-identify a person or reveal a trade secret will be prevented.

From a privacy perspective, the company’s contractual permission for the supplier to use data for additional purposes does not replace the legal basis required to process personal information relating to its customers or employees. Likewise, a contractual right to have files deleted does not ensure that information already used to train a model can be removed from that model on an individual basis. Where training on customer information is permitted, its purposes, scope and the technological limitations on subsequent deletion or separation should be addressed in advance.

Privacy, Trade Secrets and Information Security

Transferring personal information to an AI system requires an assessment of the obligations of the controller of the database and the status of the supplier as a party processing information on its behalf, insofar as those are their actual roles under Israeli law. Regulation 15 of the Protection of Privacy Regulations (Data Security), 5777-2017, requires, in the circumstances specified there, assessment of outsourcing risks and contractual arrangements governing processing purposes, permitted activities, information security, incident reporting, return of information and its destruction. A contract stating that “privacy is solely the customer’s responsibility” does not release the supplier from obligations imposed on it directly by law. Separate legal obligations also govern notices to users and the permitted purposes of processing their information.

Where the AI system receives source code, quotations, customer lists or research documents, it may also receive trade secrets. In that case, confidentiality obligations should be suited to the service, including restrictions on the use of information, commitments binding the supplier’s employees and subcontractors, access controls and a prohibition on exposing one customer’s information to another. Describing information as “confidential” in the contract does not eliminate the need to take practical measures to protect it.

The services agreement should specify how suspected security incidents are reported, how activity logs are preserved and who bears the costs of investigation and response. Prompt notification by the supplier may be essential for the customer to comply with obligations toward the Israeli Privacy Protection Authority. In an AI agreement, ordinary information-security and incident-response provisions should also address the risk of disclosure through prompts, outputs, retrieval mechanisms and tools connected to the model.

Rights in the Model, Customizations and AI-Generated Outputs

An AI agreement should distinguish between an existing foundation model, software components developed by the supplier, customizations made for the customer, the datasets used for those customizations and the outputs generated during use. Payment for implementing a system does not necessarily give the customer ownership of the model or its parameters, just as a licence to use a service does not automatically confer a right to transfer the model to another operator. Where the customer finances dedicated training or customization, rights to use, maintain, commercialize and transfer the resulting components should be expressly addressed.

For outputs, contractual permission to use a result must be distinguished from the existence of copyright protection. A supplier may permit commercial use of an output, but cannot guarantee legal exclusivity in a result in which no copyright subsists or promise that a similar result will not be generated for another user. Questions concerning protection of works created with human assistance, ownership by an employer or commissioning party, and rights in third-party content are considered separately in our guide to ownership of AI-generated content in Israel. For the present agreement, the question is which usage rights the supplier undertakes to grant and what exceptions apply.

A separate risk concerns infringement of third-party rights, for example where an output includes protected material whose use is not permitted or where a model provider used materials without appropriate permissions. Liability does not depend solely on who operated the system. The source of the material, contractual terms and the user’s conduct must be examined. Accordingly, indemnity clauses should distinguish between claims arising from technology supplied by the provider and claims arising from materials supplied by the customer or use outside the agreed instructions. Copyright infringement involving AI presents a distinct legal issue.

Performance Metrics, Acceptance Testing and Model Changes

For an AI system, a general promise that the service will be “accurate” or “reliable” is insufficient. Where the system classifies requests, produces documents or recommends actions, the parties should define the agreed use, the data on which the system will be tested, how errors will be measured and the failure rate the customer is prepared to accept. A promise that the system will never make a mistake is not a workable performance standard. Testing against an agreed set of cases, documenting error categories and establishing a correction procedure make it possible to determine whether the supplier has fulfilled its obligations.

During implementation, separate acceptance tests can be established for interface functionality, responses to specified scenarios, treatment of missing information and referral of decisions for human approval. Re-testing may also be needed after launch, because output quality may change as a model is updated, the knowledge base changes or users’ behaviour develops. The agreement should specify when material changes require advance notice, when the customer may reject a change and what remedies apply if a new version fails to meet the agreed metrics.

An AI service may contain several layers with different availability measurements: the model server, knowledge bases, APIs and mechanisms for operating connected tools. A conventional uptime metric does not reveal whether the system was technically available but produced results that failed to meet professional requirements. Commitments concerning availability and support should therefore be distinguished from those relating to output quality, security and the proper execution of actions.

Autonomous AI Agents: Authority to Act on Behalf of the Business

A system that merely provides an answer creates a different risk profile from a system authorized to send quotations, credit customers, delete records or make purchases. When an AI agent is connected to an organization’s systems, a misinterpreted instruction or manipulation through content the agent has read can turn an incorrect output into an action with financial consequences. The agreement should specify which systems the agent may operate, what data it may change, the limits on its authority and which actions require approval by an authorized person.

Human approval mechanisms should be tailored to the action concerned. A manager’s one-time approval to activate the system is not equivalent to approval of an exceptional transaction or the transfer of sensitive information. The arrangements should also permit rapid revocation of access, retention of records of instructions given to the agent and actions it performed, and reconstruction of the decision sequence after an incident. Where the supplier can update the agent’s operating instructions, the agreement should specify the conditions under which it may change the scope of the agent’s authority.

There is no contractual justification for automatically excluding all supplier liability for an action merely because it was carried out by “artificial intelligence”. Equally, the supplier cannot control every decision of a customer’s employee who grants the system unlimited permissions. Responsibility should be allocated by reference to each party’s ability to prevent the risk, including the design of permission controls, the quality of oversight mechanisms, the selection of connected applications and the system’s actual operation.

Incorrect Outputs, Indemnification and Limitations of Liability

Liability provisions in an AI agreement should distinguish between the system’s failure to comply with the agreed service description and loss caused by the customer’s unchecked reliance on a recommendation. If the supplier undertook that the system would review documents according to defined rules, the question is not only whether the result was wrong, but whether those rules were breached and whether that breach caused loss. If the customer used a tool intended to assist with drafting to make automatic decisions on eligibility for a benefit, it is necessary to examine whether that use fell within the agreed contractual purpose and what the parties knew about the way the system would be deployed.

A liability cap calculated by reference to subscription fees may be suitable for a temporary service disruption, but may not adequately reflect exposure caused by a personal-data leak, breach of confidentiality or intellectual property claim. Different caps can be established for different risks; certain breaches can be excluded from a cap; and specific indemnities can be agreed for third-party claims. A substantive negotiation should also address remedies other than monetary compensation, such as correction of a defect, cessation of unlawful processing, deletion of information and cooperation in defending a claim.

The scope of indemnification should be expressly examined where an incorrect outcome results from a combination of factors: faulty customer data, an external foundation model, a third-party component or a change made by the supplier after implementation. A blanket promise to indemnify the customer against “all AI-related loss” may prove meaningless if broad exclusions remove almost every practical use from its scope. Conversely, transferring the entire risk to the supplier does not replace the customer’s responsibility to define the application, restrict permissions and review outputs in areas under its control.

Obligations owed to a regulator or to a person whose information has been compromised do not disappear because the supplier and customer have allocated liability between themselves. They may agree how to allocate costs and handle their relationship, but cannot use their contract to change which party bears a direct statutory obligation.

AI Use Policies for Employees and Contractors

Even an organization that does not develop AI systems must decide how its staff may use them. An employee may enter a confidential quotation, a customer’s source code or an applicant’s personal details into an external system to obtain a summary or recommendation. Before doing so, the relevant terms of service and permissions granted to the provider should be examined: confidentiality provisions in the employment agreement do not ensure that the external system is subject to the same commitments.

An effective internal policy should identify the tools approved for use, the types of information that must not be entered, when approval is required for a new tool, how outputs are reviewed before delivery to a customer and who is responsible for records and removal of employee access. Where the business uses AI to make decisions concerning individuals, it should also define human involvement, procedures for correcting inaccurate information and the response to complaints. Not every use of a drafting aid requires legal approval. A policy differentiating between uses according to risk is more likely to operate in practice.

Similar arrangements are needed in agreements with service providers. A content supplier or independent developer may use AI systems without the commissioning customer’s knowledge, transferring the company’s information to those systems or creating work products whose usage rights are unclear. The agreement should specify whether such use is permitted, what restrictions apply and whether the contractor must disclose the use of AI in particular work. The commitment should address the risk the parties intend to manage rather than impose a blanket prohibition unrelated to the project’s needs.

Termination, Data Deletion and Transition to Another Provider

The right to terminate a subscription does not guarantee that the customer can continue the business process built around the system. Before contracting, the customer should establish whether original data, work histories, outputs, system instructions, templates and configuration files can be returned in a usable format. By contrast, a commercial foundation model or model weights not owned by the supplier may not be transferable. If portability is important, it should be made a procurement requirement rather than assumed to follow from a general right to export data.

Where a model has been customized using the company’s information, the data used for customization must be distinguished from the technological result produced by that process. An agreement can specify who may use the customization after termination, what must be delivered to the customer and what remains with the supplier. Where several suppliers are involved, cooperation in transition, disconnection of interfaces and revocation of access to the company’s systems should also be addressed.

Deletion obligations should cover copies retained by the provider, retention periods for backups and records, and confirmation that the required steps have been completed. If the provider was authorized to train a general-purpose model using customer data, the parties must distinguish in advance between deleting the source data and the ability or obligation to remove their influence from the model. Detailed exit arrangements enable the customer to assess the cost of migration and its actual dependence on the supplier before signing the contract.

For advice on AI agreements governed by Israeli law, AI-related commercial disputes or representation in Israel, international companies and referring foreign law firms may contact Drori Stav at office@dwo.co.il.

Have a question about the article?

You might also be interested in