Every indemnity answers a single question: what claims are covered? In AI agreements, the answer often turns on the vendor's documentation, a document you may not have read and the vendor may later rewrite. That is where liability quietly widens, and where a careful reader can close the gap.
I heard this put plainly in a recent How to Contract webinar on AI indemnification. John Pavolotsky, Partner at Stoel Rives, was reviewing an indemnity that covered the customer's use of the service in accordance with vendor's documentation. With that language, he said, the vendor “can drive a truck through it.” He highlighted how the vendor could update the documentation later to add restrictions. All that was because the clause the customer negotiated pointed to terms the vendor writes and can rewrite.
John’s point was echoed by Jade Hamilton, Senior Legal Counsel at Mariner. She emphasized that the documentation could reach anything the vendor has said or provided to the customer at any time while providing the service. These agreements include extensive referenced documents, so any one of them can shape what the indemnity actually covers.
Now, I know none of this is new. We have anchored indemnities to outside documents for a long time.
Ehat has changed is the product. With past physical and digital products, it was easier for customers to figure out what exactly the product did. They could see where it might fail. But AI behaves in ways that are still not well understood by the customer, and sometimes even by the vendor selling it. When the thing being indemnified is hard to predict, the document that defines proper use carries far more weight, and vendors have more incentive to draft that documentation in a way that protects them from as much of those unknown failure modes as possible.
All this leaves the customer at double the risk. First, the customer has operational and liability exposure to its own customers and contractors for a tool it may not fully appreciate. On top of that, the vendor can adjust its exposure in real time. That power comes from being able to edit the documentation after the contract is signed, further narrowing the scope applicable to the indemnity. This move reduces the benefit of the indemnification provision without any change to the contract itself.
This is why a phrase like "documentation as in effect from time to time" deserves a harder look in an AI deal than it does with more predictable products. This risk makes understanding the product, and the limits of what its documentation covers, one of the critical steps for contract teams.
Weighing the protection offered by an AI product indemnity means reading that document as closely as the clause itself, and accounting for the vendor's freedom to reduce its scope after the deal is signed.
