An AI licensing agreement should address the issues that standard software licenses miss: who owns the inputs, outputs and model improvements; whether the vendor may use customer data to train its models; who bears the risk of IP infringement claims based on training data or outputs; what accuracy and performance the vendor commits to; how each party will comply with privacy and AI-specific regulations; and what happens to data and models when the agreement ends. Getting these terms right protects both licensors and the businesses that deploy AI.
Businesses now license AI in many forms: API access to a foundation model, AI features embedded in SaaS products, custom models trained for a customer, and datasets licensed for training. The basic contract structure is familiar, but AI changes the risk profile. Here are the key areas to negotiate.
1. Intellectual Property Rights
Start by defining what is licensed and what each party owns:
- The model and software. The licensor usually keeps ownership of its models, algorithms and code. The license should state the permitted uses, users, territory and whether the licensee may modify, fine-tune or integrate the technology.
- Inputs. The customer should keep ownership of the data and prompts it provides.
- Outputs. Many vendors assign outputs to the customer or disclaim any ownership. Remember that purely AI-generated material may not be protected by copyright at all, so a contractual assignment may convey less than it appears. Customers who need exclusive rights should also consider confidentiality and use restrictions.
- Improvements and custom models. Address who owns fine-tuned models, model weights trained on the customer’s data, and improvements made during the relationship. If the customer is paying for a custom model, it may want ownership or an exclusive license.
2. Training on Customer Data
This is often the most heavily negotiated AI-specific term. Vendors may want to use customer inputs and outputs to train or improve their models; customers often want to prohibit it, especially for confidential, regulated or competitively sensitive data. Options include an outright prohibition, an opt-in, use limited to aggregated or de-identified data, or use only to improve the customer’s own instance. Whatever the parties agree, define the terms precisely and address data retention and deletion.
3. Data Privacy and Security
AI systems often process personal information, so the agreement must allocate responsibilities under laws such as the CCPA, the EU and UK GDPR and, for health data, HIPAA. Key provisions include:
- Service provider or processor terms that limit the vendor’s use of personal information to the services.
- Security standards, certifications and audit rights.
- Breach notification obligations and cost allocation.
- Restrictions on subprocessors and on cross-border transfers.
California’s CCPA regulations now include rules on automated decision-making technology and risk assessments, so a business using AI for significant decisions about consumers may need information from its vendor to comply. See our post on cybersecurity laws for AI startups for security obligations.
4. Liability and Indemnification
AI creates new categories of third-party claims. Allocate them explicitly:
- IP indemnity. Customers increasingly ask vendors to indemnify against claims that the model, its training data or its outputs infringe third-party rights. Several major providers now offer output indemnities, usually with conditions, such as using built-in safety filters and not knowingly prompting for infringing content.
- Customer indemnity. Vendors typically ask customers to indemnify for claims arising from customer data, prompts and prohibited uses.
- Limitation of liability. Caps tied to fees are standard, but parties often negotiate higher “super caps” or exclusions for data breaches, confidentiality breaches and indemnified IP claims.
- Dispute resolution. Specify governing law, venue and whether disputes go to arbitration.
5. Performance Metrics and Warranties
AI output is probabilistic, and vendors usually disclaim any warranty that outputs will be accurate. Customers can still negotiate meaningful commitments:
- Service levels for uptime, latency and support response, with service credits.
- Documentation of intended uses, known limitations and testing performed, including bias testing where relevant.
- Notice before material model changes or deprecation of a model version the customer relies on.
- Warranties that the vendor has the rights needed to provide the model and has complied with applicable law in developing it.
Customers remain responsible for how they use outputs, so human review of high-stakes outputs should be built into internal processes.
6. Regulatory Compliance and Acceptable Use
AI-specific laws are taking effect in phases. The EU AI Act imposes obligations on providers and deployers depending on risk level, Colorado has enacted a law on high-risk AI systems, and California has adopted laws on generative AI training data transparency and frontier model safety, among others. Contracts should require each party to comply with laws applicable to its role, allocate responsibility for required disclosures and assessments, and include an acceptable use policy that prohibits uses the vendor cannot support. Our post on regulatory compliance for AI-based products covers these rules in more detail.
7. Third-Party Dependencies
Many AI products are built on another company’s foundation model. If so, the vendor’s obligations may be limited by its upstream terms. Ask which models and providers are used, whether upstream terms flow down to you, and what happens if the upstream provider changes or ends its service.
8. Term and Termination
Define the license term, renewal and early termination rights. On termination, address return and deletion of customer data, including data in logs and fine-tuned models; transition assistance; continued rights to outputs created during the term; and which provisions survive.
The Takeaway
AI licensing agreements should cover more than traditional software terms. Focus on ownership of outputs and improvements, limits on training with your data, privacy and security, IP indemnities, realistic performance commitments, regulatory compliance, upstream dependencies and an orderly exit. Whether you are licensing AI in or out, tailor these terms to how the technology will actually be used.
Frequently Asked Questions
Who owns the output of an AI tool under a license agreement?
Ownership is set by contract, and many vendors assign outputs to the customer. However, purely AI-generated material may not be protected by copyright, so an assignment may give the customer fewer rights than it expects. Confidentiality and use restrictions can fill some of that gap.
Can an AI vendor train its models on my company’s data?
Only if the agreement allows it. Many enterprise agreements prohibit training on customer data or allow it only on an opt-in or de-identified basis. Read the definitions carefully, because terms like “service improvement” can be broader than they appear.
Should an AI vendor indemnify us for copyright claims?
It is worth negotiating. Several major AI providers offer indemnities for claims that outputs infringe third-party rights, usually subject to conditions. Check the exclusions, the cap and whether the indemnity covers training data as well as outputs.
What AI-specific laws affect licensing agreements?
The EU AI Act, Colorado’s law on high-risk AI systems, California’s CCPA rules on automated decision-making and other state laws can impose obligations on developers and deployers. Contracts should allocate compliance responsibilities based on each party’s role.
How Sutter Law Can Help
We draft and negotiate AI licenses, SaaS agreements and data licenses for both vendors and customers. See our contract review post for practical tips, learn more about our AI startup legal services, or contact us for a free consultation.
This article is general information, not legal advice. Reading it does not create an attorney-client relationship.
