AI Contracts: What to Include When You're Licensing AI Technology

Licensing AI technology is not the same as licensing standard software. The contracts that govern AI relationships need to address questions that simply did not exist a decade ago: who owns outputs generated by the model, what happens to your data after you feed it into the system, who is liable when the model produces a harmful result, and what rights you have if the underlying model is retrained or deprecated.

Most businesses signing AI licensing agreements are using contracts that were either adapted from standard software agreements or provided by the AI vendor without negotiation. Both approaches leave significant gaps. Standard software contracts were not designed for the specific risks AI creates. Vendor-provided contracts are written to protect the vendor.

Whether you are licensing AI technology from a third party, licensing your own AI to customers, or building a product on top of a foundation model, this post covers the contractual provisions that matter most and what to watch for in each one.

Scope of license: what you are actually getting

The license grant is the foundation of any AI licensing agreement, and in the AI context it needs to be more specific than in a standard software deal.

Define what is being licensed. Is the license covering access to a model via API, a deployed instance of the model, the model weights themselves, or a combination? These are meaningfully different things. API access gives you use of a model someone else controls and can modify or discontinue. Model weights give you a copy of the model you can run independently. The rights and risks attached to each are very different.

Clarify the permitted use scope. AI licenses often include restrictions on use that go beyond what software licenses typically contain. Common restrictions include prohibitions on using the model to train competing models, restrictions on the types of inputs that can be processed, and limitations on commercial use for certain tiers. Understand every restriction before you build a product or workflow around the licensed technology.

Address fine-tuning and modification rights. If you plan to fine-tune a licensed model on your own data, the agreement needs to explicitly address whether fine-tuning is permitted, who owns the fine-tuned model, and whether the licensor can access or use your fine-tuned version. Many AI licenses are silent on this point, which creates significant ambiguity about rights in a model your team has spent months improving.

Specify exclusivity if it matters. In competitive markets, an exclusive license to a particular AI capability or model can be a meaningful business advantage. If exclusivity is important to your commercial strategy, negotiate for it explicitly and define its scope carefully — geographic exclusivity, exclusivity within a particular vertical, or exclusivity for a defined period all mean different things.

Data rights: the most important and most overlooked provision

How your data is handled under an AI licensing agreement is where the most consequential risks live, and it is the section most frequently left inadequately addressed.

What happens to your inputs. When you send data to an AI system — customer queries, internal documents, proprietary datasets — that data is processed by the licensor's infrastructure. The agreement needs to specify whether the licensor can retain that data, use it to improve or retrain their models, share it with third parties, or use it for any purpose beyond delivering the service you paid for. Vendor default terms often permit broad use of inputs. Enterprise agreements can and should restrict this.

Training data restrictions. If your inputs can be used to train or fine-tune the licensor's model, you are contributing to a system that your competitors may also use. This may be acceptable depending on the context, but it should be a deliberate, informed decision, not a consequence of not reading the terms. Negotiate explicit restrictions on training use if your inputs include proprietary or sensitive information.

Output ownership. Establish clearly who owns the outputs generated by the AI system when processing your inputs. Many AI vendors assign output ownership to the customer, but some retain rights or impose limitations on how outputs can be used commercially. If you are building a product on top of AI-generated outputs, you need contractual clarity that you own or have sufficient rights to those outputs for your intended use.

Data security and breach notification. Specify the security standards the licensor must maintain for your data, the encryption requirements for data in transit and at rest, and the notification timeline and obligations if a breach occurs. GDPR, CCPA, HIPAA, and other regulatory frameworks may impose specific requirements that need to be reflected in the contract.

Data deletion. Define what happens to your data when the agreement ends. The contract should require the licensor to delete or return your data within a specified period following termination and certify that deletion has occurred. Without this provision, your proprietary data may remain on the licensor's systems indefinitely after the relationship ends.

Intellectual property ownership: inputs, outputs, and fine-tuned models

IP ownership in AI agreements is genuinely more complex than in standard software deals, and the contract needs to address it with corresponding specificity.

Your inputs remain yours. The agreement should confirm that you retain all rights in the data, content, and information you provide as inputs. This sounds obvious but is worth stating explicitly because some vendor terms include broad license grants that could be read to allow the vendor to use your inputs in ways that affect your IP position.

Outputs: establish your rights clearly. As discussed in the data section, output ownership should be explicitly assigned. Given the unsettled state of AI copyright law, contractual assignment of output rights from the licensor to the licensee is the most reliable way to establish a clear chain of ownership, even in jurisdictions where copyright in AI outputs is uncertain.

Fine-tuned models: negotiate ownership before you invest. If you fine-tune a licensed model on your data, the resulting model is a hybrid: it incorporates the licensed foundation model and your proprietary improvements. Who owns that hybrid depends entirely on what the agreement says, and if the agreement is silent, you may have invested significant resources in a model the licensor can claim rights to. Negotiate ownership of fine-tuned models, or at minimum a perpetual license to use and deploy the fine-tuned version, before beginning fine-tuning work.

Background IP and derivatives. Define which party owns background IP brought into the relationship and what rights, if any, the other party has to it. Establish how derivative works created during the relationship are treated. These provisions prevent disputes about IP that was developed before the agreement from contaminating the relationship.

Liability and indemnification: who is responsible when things go wrong

AI systems fail in ways that standard software does not. They hallucinate, produce biased outputs, generate harmful content, and make errors that can have real consequences. The liability provisions in an AI licensing agreement need to address these risks specifically.

Indemnification for IP claims. The licensor should indemnify you for third-party IP claims arising from the model itself — for example, if the model's training data included copyrighted content and a rights holder asserts a claim based on an output. This is a meaningful risk given the volume of training data copyright litigation currently proceeding. Understand the scope of this indemnification, the caps on it, and the conditions that must be met to trigger it.

Liability for harmful outputs. Who is responsible if the AI system generates output that causes harm — defamatory content, dangerous instructions, discriminatory decisions? Vendors typically limit their liability significantly for output-related claims. Understand those limitations and evaluate whether they are acceptable given your use case. If you are deploying the AI in a high-stakes context, you may need insurance or contractual protections beyond what the vendor is willing to provide.

Limitation of liability caps. Standard limitation of liability clauses cap the licensor's total liability at the fees paid in the preceding 12 months. For AI systems where a failure could cause significant downstream harm, this cap may be inadequate. Negotiate for higher caps or carve-outs for specific categories of harm if your deployment context warrants it.

Regulatory compliance allocation. Specify which party is responsible for compliance with applicable regulations, including the EU AI Act, FTC requirements, EEOC guidance on AI in employment, and sector-specific regulations. Do not leave this implicit. If your use of the AI system triggers regulatory obligations, the contract should clearly allocate responsibility for meeting them.

Model performance, availability, and deprecation

These provisions are frequently overlooked in AI agreements and consistently become sources of dispute.

Performance standards. Unlike standard software with defined features, AI model performance is probabilistic and can change. The contract should specify what performance standards the model is expected to meet, how performance is measured, and what remedies are available if the model underperforms. Be specific: accuracy thresholds, latency requirements, and availability standards should all be defined.

Model updates and versioning. Licensors update their models regularly, and updates can degrade performance on your specific use case. The agreement should address whether you have the right to remain on a specific model version, how much notice you will receive before the model is updated in ways that could affect your product, and what testing rights you have before a new version is deployed in your environment.

Deprecation and continuity. AI models get deprecated. Foundation models get replaced by newer versions. The contract should specify the minimum notice period before a model you depend on is deprecated, whether the licensor will provide transition assistance, and what access rights you retain to model weights or outputs after deprecation if you need them for continuity.

Audit rights. For AI systems used in regulated contexts or where performance monitoring is important, negotiate the right to audit the system's outputs and the licensor's compliance with the agreement's terms. Audit rights are increasingly standard in enterprise AI agreements and should be expected.

Termination provisions

Termination for cause and convenience. The agreement should specify the circumstances under which either party can terminate, the notice required, and what happens to data, outputs, and fine-tuned models upon termination.

Effects of termination. Define precisely what each party's obligations are when the agreement ends. Data deletion timelines, return of proprietary materials, survival of confidentiality obligations, and post-termination use rights all need to be addressed. Without clear termination provisions, disputes about what each party can do after the relationship ends are almost inevitable.

Business continuity. If your product depends on the licensed AI, termination of the agreement could be catastrophic. Consider negotiating source code escrow arrangements or model weight escrow for critical AI dependencies so that your product can continue operating if the licensor becomes unable to perform.

Frequently asked questions

Do I need a lawyer to review an AI licensing agreement?

Yes, particularly if you are building a commercial product on the licensed technology, sharing proprietary data with the licensor, or entering a high-value or long-term relationship. Standard vendor terms are written to protect the vendor. The provisions that matter most for your protection — data rights, IP ownership, indemnification, deprecation — are often the ones that receive the least favorable treatment in standard terms.

What is the difference between an API license and a model license?

An API license gives you the right to access and use a model via the licensor's infrastructure. You do not receive the model itself, and the licensor can update, modify, or discontinue the model. A model license grants you rights in the model itself — potentially including the weights — which you can deploy independently. Model licenses give you more control and continuity but typically cost significantly more and involve more complex IP negotiations.

Can I use outputs from a licensed AI model in my commercial product?

It depends on the license terms. Many AI licenses explicitly permit commercial use of outputs. Some impose restrictions, particularly for lower-tier or free-tier access. Review the vendor's terms carefully and ensure that the agreement explicitly addresses commercial use of outputs for your specific intended use case.

What should I look for in an AI vendor's standard terms before I sign?

Focus on four areas: the data use provisions (can the vendor use your inputs to train their model), the output ownership provisions (who owns what the model generates), the indemnification scope (are you protected against IP claims arising from the model), and the limitation of liability (is the cap appropriate for your risk exposure). If any of these are inadequate for your use case, negotiate before signing.

How should AI licensing agreements handle regulatory requirements?

The agreement should explicitly allocate regulatory compliance responsibilities between the parties. For US companies, this includes FTC requirements, sector-specific regulations, and state AI laws. For companies with EU operations or customers, it includes the EU AI Act. The allocation should reflect which party is best positioned to control the relevant compliance factors — typically the licensor for model-level requirements and the licensee for deployment-level requirements.

AI licensing agreements are among the most consequential contracts your company will sign, and the standard forms currently circulating in the market were not designed with your interests as the primary concern. The provisions that protect your data, your IP, your liability position, and your business continuity require active negotiation.

If you are entering an AI licensing relationship and want counsel to review or negotiate the agreement, contact Ana Law to schedule a strategy session.