Your AI Vendor Contract Already Decided Who Pays

Picture the closing table. The file is clean on its face. Everyone is present. One number on one line came out of a tool.

The number is wrong. Not maliciously wrong. Wrong the way a prediction engine is wrong, confidently and in the correct format.

The reflex question is who pays. The reflex answer is the vendor.

The vendor already answered that question, in a contract nobody at your firm has read.

---

Forensic Case File

What follows is an illustrative scenario. It is not a real matter, a real firm, or a real client.

Thursday afternoon. A payoff figure comes back from an AI-assisted title tool your firm licensed last spring. It reads clean. The number is close enough to plausible that nobody double-checks it against the underlying file.

The wire goes out. Six weeks later, the servicer flags a discrepancy. The tool produced a number that was internally consistent but wrong.

Someone pulls the vendor agreement for the first time since it was signed. Buried in Section 14 is a clause disclaiming any warranty as to the accuracy of output and shifting reliance risk to the user. Nobody at the firm had read it. The vendor was never required to explain it.

The firm covers the difference. The disclaimer holds.

What Does Your AI Vendor Contract Actually Say About a Bad Output?

Two separate things are true, and they come from two different places.

The first comes from the State Bar of Georgia. Its Generative AI Toolkit tells Georgia lawyers that standard-form clickwrap and shrinkwrap agreements, the kind accepted by using the product, frequently include broad disclaimers and may expressly exclude any obligation of confidentiality. In its confidentiality analysis, the Toolkit goes further. It states that terms of service commonly allocate the risk inherent in the technology to the user, with the vendor disclaiming responsibility. The Bar put that in writing, to Georgia lawyers, about the contracts they are signing.

The second comes from reading the contracts themselves. Most AI vendor agreements disclaim any warranty as to the accuracy of output. That is not a Bar finding. It is what the documents say, and it is why the Toolkit tells you to read them.

Put together: the Bar told you these agreements shift risk and carry broad disclaimers. Your specific contract tells you exactly how far. Almost nobody has looked.

The Toolkit includes a contract review checklist for this reason. It asks firms to examine the scope of client data use, whether the vendor trains its models on identifiable client data, whether obligations flow down to the underlying model provider, what assurances exist against biased output, and what happens to prompts and output when the contract term ends.

Most firms using an AI tool today have never run that checklist against their own vendor.

Which Clauses in Your AI Vendor Contract Shift the Risk to You?

Three of them. The warranty disclaimer, the limitation of liability, and the indemnification provision.

They sit in different parts of the agreement, they are rarely labeled in plain terms, and together they do most of the work.

The warranty disclaimer. Usually set in capital letters somewhere past the middle of the document. It states that the service is provided as is, without any warranty of accuracy, reliability, or fitness for a particular purpose. Read plainly, the vendor is telling you it does not promise the output is correct. Everything the marketing page said about accuracy sits outside the contract. This clause is the one that governs.

The limitation of liability. This does two things at once. It caps the vendor’s total exposure, often at the fees you paid in the preceding twelve months, and it excludes consequential and indirect damages entirely. A wire sent on a wrong number is a consequential damage. If your annual subscription is a few thousand dollars, that is the ceiling on what the vendor could owe, and the category of loss you actually suffered is likely carved out above it.

The indemnification provis