Most California businesses do not build artificial intelligence—they buy it. They license a chatbot, an applicant-screening platform, a pricing engine, or a generative-content tool from a vendor and bolt it into their operations. When that tool fails, infringes someone's copyright, leaks customer data, or produces a discriminatory result, the contract with the vendor is the single most important document determining who absorbs the loss. Yet most AI agreements arrive as vendor-friendly click-through terms engineered to push every risk onto the customer.

This guide covers the clauses California businesses should negotiate before signing—indemnification, intellectual property, data and confidentiality, performance, regulatory compliance, limitation of liability, audit rights, and termination—so the contract protects your company instead of the company that sold you the software. If a vendor relationship has already gone wrong, our team handles business contract disputes throughout California.

Why an AI Contract Is Not Like an Ordinary Software Contract

Traditional software does what it is programmed to do. AI tools fail in distinctive and unpredictable ways: they “hallucinate” false statements, reproduce bias from their training data, generate content that infringes third-party intellectual property, and change behavior over time as the underlying model is updated. They also ingest your data—sometimes to train the vendor's general model.

This risk profile collides directly with California's new liability landscape. Under AB 316 (Civ. Code § 1714.46), effective January 1, 2026, your business cannot defend an AI-related claim by arguing the AI acted autonomously. If you cannot disclaim responsibility to the injured party, your contract with the vendor becomes the primary mechanism for shifting that risk to the party that actually developed and trained the system. (For more, see our guide to AI liability for businesses.)

Indemnification: The Clause That Matters Most

Indemnification decides who pays when a third party sues. In an AI contract, insist that the vendor indemnify your business against claims that the tool's outputs or training data infringe intellectual property, against losses from a data breach caused by the vendor, and against third-party claims arising from defects in the vendor's system.

Read the indemnity for the traps that gut it: a low dollar cap, a carve-out for “modified” or “combined” use that swallows the protection, and “sole and exclusive remedy” language that bars every other claim. For infringement arising from training data—a live risk in current AI litigation—push hard for an uncapped indemnity.

Intellectual Property: Who Owns the Inputs and the Outputs

An AI contract should state clearly who owns three different things: the data you feed in, the outputs the tool generates, and any custom or fine-tuned model created from your data. Negotiate ownership or, at minimum, a broad commercial license to use the outputs. Prohibit the vendor from using your confidential inputs to train its general-purpose model unless you expressly agree. And recognize a limitation of the law: purely AI-generated material may not qualify for copyright protection, so do not assume an output is a protectable asset. Where proprietary material is at stake, our intellectual property practice can help structure the terms.

Data, Privacy, and Confidentiality

If the tool touches personal or confidential information, the contract needs a data processing addendum and clear answers to basic questions: Where is the data stored and processed? Is it used for training? Who are the subprocessors? What security standards apply? Will data be returned and deleted on termination, and how fast must the vendor notify you of a breach? For any tool handling California consumers' personal information, the agreement must align with the CCPA/CPRA, including the contractual terms those statutes require.

Performance, Accuracy, and Service Levels

Vendors love to disclaim all warranties of accuracy. Resist. Negotiate representations about accuracy and bias testing, meaningful service-level commitments, and a defined remedy if the model degrades or materially changes after deployment. Build in acceptance testing before go-live, and reserve the right to require human review of consequential outputs—both to manage risk and to satisfy the “reasonable oversight” expectations that drive AI liability after AB 316.

Compliance With Law: Allocating Regulatory Risk

California's AI rulebook is expanding quickly. Require the vendor to represent that its tool complies with applicable law—including anti-discrimination statutes, AI transparency and disclosure requirements, and, for pricing tools, the new algorithmic-pricing restrictions under AB 325. Add an affirmative obligation to update the tool as the law changes, and specify which party is responsible for any legally required customer disclosures.

Limitation of Liability: Read the Cap Before You Sign

The limitation-of-liability clause is where vendors quietly cap their exposure at the fees you paid—often a trivial sum next to the harm a failed AI system can cause. Negotiate a higher cap, and carve the categories that matter most out of the cap entirely: indemnification obligations, breaches of confidentiality, data-security incidents, and IP infringement. Push for the limitation to run both ways.

Audit Rights, Records, and Explainability

Because liability now turns on whether your deployment was reasonable, you need access to proof. Negotiate the right to receive documentation about the vendor's training, testing, and bias-mitigation practices, the right to audit or obtain a third-party assessment, and—for tools that make consequential decisions—a degree of explainability you can stand behind if a regulator or plaintiff asks how the system reached its result.

Termination and Transition

Finally, secure an exit. Negotiate the right to terminate for a material adverse change to the model, a security incident, or a failure to maintain legal compliance—not just for non-payment. On the way out, require prompt return and deletion of your data, reasonable transition assistance, and terms that prevent the vendor from holding your operations hostage through lock-in.

Frequently Asked Questions

Q: We are signing a standard click-through AI subscription. Can we even negotiate?
A: Leverage varies with your spend, but you can almost always address the highest-risk items by addendum—indemnification, a prohibition on using your data for training, liability carve-outs, and a data processing addendum. Even modest customers can often secure these.

Q: If the vendor indemnifies us, are we protected from AB 316 liability?
A: Indemnity shifts the cost of a loss between you and the vendor, but it does not make you invisible to the plaintiff. You remain liable to the injured party; the contract is how you recover that loss from the vendor. That is precisely why the indemnity language matters so much.

Q: Who owns what the AI produces for our business?
A: It depends entirely on the contract. Negotiate ownership or a broad commercial license to the outputs, and remember that purely AI-generated material may not be protectable by copyright.

Q: Can the vendor use our confidential data to train its model?
A: Only if your contract allows it. Prohibit or tightly restrict training use of your inputs, especially for confidential, proprietary, or personal data.

Q: A sales representative promised the tool was “fully compliant.” Is that enough?
A: No. Oral assurances rarely survive a contract's integration clause, which typically bars reliance on promises outside the written agreement. If it matters, get it in the contract as a representation or warranty.

This article is provided for general informational purposes and is not legal advice.

Need help? Contact Kolmogorov Law, P.C. at (909) 235-6116 or visit kolmogorovlaw.com to schedule a consultation with our business litigation team in Irvine, California.