Your AI Vendor Wants Your Data but None of the Liability

AI vendor agreements often resemble ordinary software contracts. But the technology introduces risks that traditional agreements may not address, particularly when it comes to customer data, generated content, and changes to the underlying model.

For instance, a $2,000 annual software subscription can come with a $2,000 liability cap. But if the software exposes confidential client records or generates material that infringes someone else's copyright, that cap starts to look considerably less reasonable.

Yet businesses routinely accept these terms without much thought. Someone signs up for an AI platform, clicks through the agreement, and starts using it. The contract receives attention only after something goes wrong.

1. What Happens to the Information You Upload?

An employee uploads confidential client documents to a popular generative AI tool for summarization. Can the vendor use those documents to train its model? Can it retain the prompts or share information with another provider?

The answer depends on the agreement. Watch for broad definitions of "usage data," "interaction data," and "service improvement." Those provisions may give the vendor rights extending beyond what the customer intended.

A promise not to sell customer data is not the same as a promise not to use it for training. If confidentiality matters, the contract should address permitted uses, training restrictions, retention, and deletion expressly.

2. Who Owns the Output, and Who Defends You If It Infringes

If an AI tool generates text, code, or images resembling someone else’s copyrighted work, the business using the output may be sued along with the vendor.

To address this concern, some large vendors have offered to absorb that risk. In 2023, for example, Microsoft announced that it would defend commercial customers of its Copilot services against copyright claims arising from their use of Copilot output. However, the commitment came with conditions: customers had to use the built-in guardrails and content filters and could not input material they lacked the right to use. But not every vendor offers comparable protection; even when an indemnity exists, its exclusions may determine whether it provides any meaningful benefit.

Another complication is that the output itself may not be copyrightable. Under U.S. law, copyright protects only human authorship. Consistent with that rule, the U.S. Copyright Office concluded in January 2025 that material generated solely by AI generally lacks copyright protection. However, human creative input, including sufficiently original editing or arrangement, may qualify. The D.C. Circuit reached the same conclusion in Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), and the Supreme Court declined review in March 2026.

In short, a vendor can assign whatever rights it has in the output, but it cannot create a copyright the law does not recognize. Therefore, if ownership of the finished work matters, it is important to document the human editing and arrangement that went into it.

3. “As Is” Means What It Says

Most AI agreements disclaim warranties that outputs will be accurate, reliable, or suitable for a particular purpose. That may be manageable when employees use AI to brainstorm ideas. However, it becomes more consequential when the software generates financial calculations, technical specifications, or materials sent directly to customers.

If the tool makes a costly mistake, the vendor may argue that checking its accuracy was your responsibility. For business-critical applications, consider whether the vendor will agree to measurable performance commitments, correction obligations, or other contractual remedies.

4. How Much Can You Recover?

Most AI agreements cap the vendor’s total liability. Usually, however, the more important question is which claims are excluded from the cap.

Illinois courts generally enforce limitation-of-liability provisions in commercial contracts, reflecting a strong public policy favoring freedom of contract. Waivers of consequential damages, however, are strictly construed against the party relying on them, so the wording can change the result.

Westlake Financial Group, Inc. v. CDH-Delnor Health System, 2015 IL App (2d) 140589, illustrates the point. There, the contract excluded “indirect” and “consequential” damages, and the Illinois Appellate Court held that the clause barred consequential lost profits but not lost profits resulting directly from the breach.

The cap also matters when compared with the vendor’s other promises. If the IP indemnity and data security obligations sit under the same low cap as everything else, those promises may be worth less than they appear.

5. Who Is Indemnifying Whom?

Indemnification provisions determine who pays when a third party sues. Customers often assume the vendor will protect them from claims caused by the vendor’s system. But some AI terms do the opposite: they require the customer to indemnify the vendor for claims arising from output, even output generated by the vendor’s own model. For that reason, the provision deserves a close enough read to know who is protecting whom.

6. The Model You Tested May Not Be the Model You Use

Vendors frequently update, retrain, and replace AI models. As a result, a tool that performed well during a trial may behave differently a few months later, and standard terms may allow those changes without notice.

Useful protections include advance notice of material model changes, a definition of “material,” and a right to terminate if performance materially declines. These usually need to be added specifically, because traditional termination provisions address breach and insolvency, and many do not contemplate model deprecation, degraded performance, or substitution of the underlying model.

7. Who Else Has Access to Your Data?

Many AI products run on a foundation model supplied by another company, so customer data may move through more than one business. Because a customer’s protections are only as strong as the least careful company in the chain, it helps to obtain a current list of subprocessors, notice of changes to that list, and a firm deadline for reporting security incidents.

8. What Happens When You Leave?

Termination is easy to overlook when the product is working.

But consider what happens if your business needs to switch vendors. Can it export its prompts, outputs, and uploaded documents? What happens to information retained in backups or used to customize a model?

Deletion provisions should identify the information covered, provide workable deadlines, address backups, and specify whether the vendor must certify compliance. Otherwise, ending the subscription may not end the vendor's possession or permitted use of your information.

Read the Terms While They Can Still Be Changed

Many AI agreements are drafted to protect the vendor, and the default terms often leave the customer with most of the risk. Although smaller purchases may offer little room to change the terms, knowing where the risk sits still lets a business plan around it by insuring, limiting how the tool is used, or choosing another vendor.

Once a problem arises, the contract is no longer a draft to negotiate. It is the rulebook for deciding who pays.

Before your business signs an AI vendor agreement, we can review the terms and identify where the risk sits. Contact us to discuss your needs.

This article provides general information and does not constitute legal advice. For guidance concerning a specific agreement, see our contract review and advisory services. Contractual rights and remedies depend on the agreement, relevant circumstances, and applicable law. To discuss potential representation, schedule a consultation. Submitting an inquiry does not establish an attorney-client relationship.