# Glossary: who pays when AI work goes wrong Canonical HTML: https://spoolis.com/guides/glossary. This page is also available in machine-readable Markdown. Plain definitions of the contract and insurance terms that decide who pays when an AI agent makes a costly mistake: warranties, guarantees, limitation of liability, E&O, exclusions and more. ## Terms **AI agent.** Software built around a generative AI model that can take actions on its own, such as filing, paying or submitting, rather than only answering questions. NIST describes AI agent systems as "at least one generative AI model and scaffolding software that equips the model with tools to take a range of discretionary actions." Source: [NIST CAISI, request for information on AI agent security (Federal Register 2026-00206)](https://www.federalregister.gov/d/2026-00206). **Consequential AI work.** Our term for work an AI agent does where a mistake brings a bill from someone outside the deal, such as a tax penalty, a payment sent to the wrong account or a denied claim. **AI agent guarantee.** Our term for a promise that if a specific piece of AI work is wrong and the mistake costs the customer money, the vendor stands behind it. **AI performance warranty.** A promise that an AI system will meet a stated performance level, such as an accuracy rate, over a period. It typically pays when the system misses the agreed number, not when one particular job goes wrong. **Warranty (express).** A seller's statement of fact or promise about what it sells that becomes part of the deal. Under the Uniform Commercial Code, it creates a warranty that the goods will match the statement. Source: [UCC § 2-313 (Cornell LII)](https://www.law.cornell.edu/ucc/2/2-313). **Warranty (implied).** A warranty the law adds to a sale even if the seller says nothing, such as the promise that goods are fit for their ordinary purpose ("merchantable"). Sellers can often exclude it in writing. Source: [UCC § 2-314 (Cornell LII)](https://www.law.cornell.edu/ucc/2/2-314). **Guarantee (guaranty).** In law, a promise to answer for someone else's obligation if they fail to meet it. In marketing, a promise of a remedy, such as a money-back guarantee. The FTC says a seller should call something a money-back guarantee only if it refunds the full purchase price on request. Source: [FTC Guides, 16 CFR § 239.3](https://www.ecfr.gov/current/title-16/chapter-I/subchapter-B/part-239/section-239.3). **Insurance.** A contract in which one party promises to pay another for a loss caused by an uncertain event. California defines it as a contract "whereby one undertakes to indemnify another against loss, damage, or liability arising from a contingent or unknown event." Source: [California Insurance Code § 22](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=INS§ionNum=22). **Surety.** A party that becomes answerable to someone for another party's performance, as with a performance bond. Source: [California Department of Insurance glossary](https://www.insurance.ca.gov/01-consumers/105-type/95-guides/20-Glossary/). **Limitation of liability.** A contract term that caps what one side must pay the other if something goes wrong, often at the fees paid over a period, and often excluding indirect losses entirely. Under the UCC, a limited remedy that "fail[s] of its essential purpose" can be set aside. Source: [UCC § 2-719 (Cornell LII)](https://www.law.cornell.edu/ucc/2/2-719). **Direct (general) damages.** Losses that flow naturally and directly from a breach. New York's highest court describes them as "the natural and probable consequence of the breach." Source: [Biotronik A.G. v. Conor Medsystems Ireland, Ltd., 22 N.Y.3d 799 (2014)](https://caselaw.findlaw.com/court/ny-court-of-appeals/1661662.html). **Consequential damages.** Losses that come from a breach indirectly, because of the buyer's particular situation, such as penalties or lost business. Contracts often exclude them, and whether a given loss counts as direct or consequential can decide whether anything is paid. Source: [UCC § 2-715 (Cornell LII)](https://www.law.cornell.edu/ucc/2/2-715). **Indemnification.** A promise by one party to cover another party's losses or costs from a specified kind of event, such as a third-party lawsuit. Source: [Cornell LII Wex: indemnity](https://www.law.cornell.edu/wex/indemnity). **Service level agreement (SLA).** A provider's commitment to a customer about the service, such as uptime or response time, and what happens when it falls short. Source: [NIST CSRC glossary: service level agreement](https://csrc.nist.gov/glossary/term/service_level_agreement). **Service credit.** The usual remedy in an SLA: a credit against future fees when the provider misses a target. It rarely covers what the shortfall cost the customer. **Errors and omissions (E&O) insurance.** Insurance that protects a business against liability for mistakes or failures in the professional services it provides. Also called professional liability insurance. Source: [Insurance Information Institute: professional liability insurance](https://www.iii.org/article/professional-liability-insurance). **Exclusion.** A part of an insurance policy that lists causes or conditions the policy does not cover. Source: [California Department of Insurance glossary](https://www.insurance.ca.gov/01-consumers/105-type/95-guides/20-Glossary/). **Endorsement.** A form attached to an insurance policy that adds, removes or changes coverage. Also called a rider. Source: [NAIC: what is an insurance endorsement or rider?](https://content.naic.org/article/consumer-insight-what-insurance-endorsement-or-rider). **First-party and third-party coverage.** First-party coverage pays the policyholder for its own losses. Third-party coverage pays others who were harmed by the policyholder, and usually only what the policyholder is legally liable for. **Silent AI.** Insurance policies that neither clearly cover nor clearly exclude losses involving AI, leaving the answer to be argued after a loss. The term echoes "silent cyber."