AI Indemnity and Model-Risk Clauses in Contracts (2026)

What these clauses are for When a business adopts an AI tool, two new risks arrive that older contracts do not address well. The first is intellectual property risk: an output, or the tool itself, might infringe someone's copyright, patent or trade mark, often because of what the model was trained on. The second is model risk: the tool might produce wrong, biased or fabricated results that cause loss when someone relies on them. AI indemnity and model-risk clauses exist to decide, in advance, who carries the financial consequences when either risk materialises. An indemnity is a promise by one party to cover the other's defined losses, usually including legal costs, if a stated event happens. A model-risk clause is broader: it combines disclosure of limitations, allocation of liability, and often caps and carve-outs, to set how far a vendor stands behind the quality of its model. Used together they turn a vague worry into a priced, allocated risk. This page is general information, not legal advice. Indemnities and liability caps are heavily shaped by local law and by the bargaining power of each side, so have a qualified professional adapt the wording for a live deal. ## The IP infringement indemnity This is the clause most AI buyers focus on, and several large vendors now offer a version of it. Adapt wording like this: ```

IP INFRINGEMENT INDEMNITY The Vendor will defend the Customer against any third-party claim that the Service, or an Output generated by the Customer's permitted use of the Service, infringes that third party's [copyright / patent / trade mark / trade secret], and will indemnify the Customer against damages and reasonable costs finally awarded or agreed in settlement, provided that: (a) the Customer promptly notifies the Vendor of the claim; (b) the Customer did not modify the Output or use it outside the documentation or in breach of this agreement; (c) the claim does not arise from Customer inputs or Customer data; and (d) the Vendor has control of the defence and settlement. This indemnity is [subject to / excluded from] the liability cap in clause [X]. Read it carefully, because the conditions are where the protection narrows. Common carve-outs exclude claims caused by your own inputs, by modifying the output, by using the tool outside its documentation, or by combining it with other software. A vendor may also require you to use the latest version or to have enabled certain filters. An indemnity that is capped low, or carved out for the exact way you use the tool, offers less than it appears. It is also worth checking what the vendor controls when a claim is made. Most indemnities give the vendor control of the defence and settlement, which is normal, but you want the right to be consulted, a say if a proposed settlement admits wrongdoing that affects you, and a clear picture of what the vendor may do to the service if a claim succeeds: whether it can modify the tool, procure a licence, or simply withdraw the feature and refund you. A modify-or-withdraw remedy can leave you without the capability you relied on, so for a tool embedded in your operations press for a continuity obligation or a reasonable transition period rather than an abrupt switch-off. ## Model-risk and accuracy clauses No credible vendor warrants that an AI model is always correct, and you should be suspicious of one that does. Realistic model-risk clauses do three things: **Disclose limitations.** The vendor states the intended use, the known limitations, and that outputs may be inaccurate or incomplete and should be reviewed. Sample: MODEL LIMITATIONS AND RELIANCE The Customer acknowledges that the Service uses probabilistic models that can produce inaccurate, incomplete or unexpected Outputs. The Vendor makes no warranty that any Output is accurate, fit for a particular purpose, or free from bias. The Customer is responsible for human review of Outputs before relying on them for any decision with [legal, financial, safety or similar] effect.

- **A super-cap or uncapped carve-out** for data breach, IP infringement, or the vendor's wilful misconduct, versus a single low cap for everything.
- **Mutual versus one-way** indemnities. You will often indemnify the vendor for your own inputs and misuse; make sure that is proportionate.
- **Exclusions of indirect and consequential loss**, which in an AI context can swallow the very losses you care about, so define the important direct losses carefully. ## What these clauses cannot do Be honest about the limits, because the biggest mistakes come from treating a clause as a safeguard: - An indemnity does not stop infringement or a harmful output happening; it decides who pays afterwards, within the cap.
- No clause makes a model accurate or unbiased. The real controls are the vendor's disclosure, your human oversight, and testing before deployment.
- An indemnity is only as good as the vendor's ability to pay. For a small vendor, a large promise may be worth little; consider insurance requirements.
- Caps and carve-outs can quietly return the risk to you. The protection lives in the detail, not the heading. ## Keeping liability terms consistent across your contracts The hard part is not writing one good indemnity; it is making sure every AI contract carries an equivalent one, that caps and carve-outs do not drift from vendor to vendor, and that when a claim lands you can quickly find which version each supplier signed and what it actually promised. Drafting from a clean clause set, spotting where a cap has crept inside the general limit or a carve-out has reappeared, editing a clause in plain language across several contracts, and holding every version with its date can live in one AI workspace tuned to how your company contracts. [Adira](https://adiralaw.com) is built around that single-workspace idea for business contracts. For a single deal the wording above is enough to start from; the point worth keeping is that an indemnity only protects you if it is consistent, uncapped where it matters, and easy to find when you need it.