Your AI Feature Made a Warranty and Your Contract Says AS IS. One of Them Loses.
Warranty law does not care what you called it. An affirmation of fact or a description that becomes part of the basis of the bargain creates an express warranty, with no magic words and no intention to warrant required. AI go-to-market is built almost entirely out of affirmations of fact — accuracy percentages, benchmark charts, demos that complete flawlessly, sales emails confirming the model handles a prospect's exact use case — while the contract underneath disclaims everything in a single generic paragraph. Those two documents are in conflict, and the specific one usually wins.
How a Promise Gets Made Without Anyone Deciding to Make One
The commercial law framework most people vaguely remember comes from Article 2 of the Uniform Commercial Code, and its treatment of express warranties is deliberately broad. Any affirmation of fact or promise relating to the goods that becomes part of the basis of the bargain creates a warranty that the goods will conform to it, as does any description of the goods and any sample or model. The drafters added that the seller need not use formal words like "warrant" or "guarantee" and need not have any specific intention to make a warranty. Puffery — vague enthusiasm about how great the product is — stays outside. Specific, verifiable claims do not.
Now consider how AI features are actually sold. The distinguishing move of AI marketing is specificity, because specificity is what makes a probabilistic system sound trustworthy: a percentage, a benchmark, a named document type, a claim about hours saved. Meanwhile "sample or model" maps almost too neatly onto a sales demo. The typical enterprise AI sale generates several express warranties before anyone opens the contract, and the people generating them are not reading it.
Where the Promises Come From
- •Unqualified accuracy percentages with no stated task, dataset or configuration
- •Benchmark charts reproduced without the conditions that produced them
- •Comparison tables asserting specific capabilities competitors lack
- •Case-study numbers presented as typical rather than as one customer's result
- •Landing-page copy promising the feature 'handles' a named document or workflow type
- •Demos run on curated inputs and presented as representative
- •Email confirmations that the model works on the prospect's specific data
- •Security and compliance questionnaire answers overstated to clear the gate
- •Pilot success criteria agreed verbally and never written into the contract
- •Roadmap commitments described in the present tense
- •Confidence scores presented as reliability without a stated meaning
- •In-product copy claiming review, verification or validation that does not occur
- •The assistant itself asserting capabilities, coverage or guarantees
- •Documentation describing behavior the current model no longer exhibits
- •Status and quality dashboards that measure something narrower than they imply
- •A single generic AS IS paragraph expected to cover every specific claim
- •No entire-agreement or no-reliance clause tying the deal to the written contract
- •No process linking marketing claims back to a measurement anyone ran
- •No right reserved to change underlying models, and no notice commitment
- •Liability caps drafted for a hosting product, never revisited for AI output
Why the Disclaimer Is Weaker Than It Reads
Disclaimers do real work, but not the work most teams assume. They are aimed primarily at implied warranties — merchantability and fitness for a particular purpose — and even that is subject to rules about conspicuousness and, in places, required language. Express warranties are treated differently: the framework instructs that warranty language and disclaimer language be read consistently where possible, and that the disclaimer yields where the two cannot be reconciled. A specific promise made to induce the purchase and a blanket denial buried in terms cannot be reconciled.
There is a second problem the contract cannot solve at all. A disclaimer is a defense to a contract claim between you and your customer. It is not a defense to a misrepresentation or unfair-practices theory, and consumer protection statutes are generally not waivable by the contract that contains the misrepresentation. That is why "we disclaimed everything" is a weaker position than it feels: it addresses the cheapest claim while leaving the more dangerous ones untouched.
The Dependency Nobody Drafted For
Traditional software contracts assume the vendor controls the software. An AI feature built on a third-party model does not fit that assumption. The provider can deprecate a version, change default behavior, alter safety filtering or shift latency and cost, and your product's behavior changes with it while your codebase sits untouched. If your agreement promised specific behavior, you now owe an obligation you do not fully control. The mitigations are unglamorous and effective: describe the feature by capability rather than by named model, reserve the right to change models and versions, commit to notice plus a testing window for material behavior changes, and put regression testing on your critical paths so you learn about drift before your customer does.
Aligning the Promise With the Product
Fix What You Say
- ☐Scope every performance number to task, dataset, configuration and measurement date
- ☐Route accuracy and compliance claims through one reviewer before publication
- ☐Run demos on realistic inputs, and say plainly what the demo data is
- ☐Keep a dated file of the measurement behind every public claim
- ☐Audit in-product copy and assistant responses for capability claims nobody approved
Fix What You Sign
- ☐Warrant something real and narrow rather than disclaiming everything unconvincingly
- ☐Separate availability commitments from output-quality commitments
- ☐If you commit to quality, define the metric, the sample and the bounded remedy
- ☐Reserve model-change rights with a notice and testing window for material changes
- ☐Include entire-agreement and no-reliance terms, and make sure sales knows what they mean
The counterintuitive conclusion is that warranting a narrow, measured, genuinely true thing is safer than disclaiming everything. A blanket disclaimer paired with aggressive marketing gives a buyer both a promise to rely on and a reason to feel misled. A modest, specific warranty that matches what your marketing says gives them a defined remedy and takes the misrepresentation story off the table.
Frequently Asked Questions
Is 'AI may produce inaccurate results' enough of a warning?
It is necessary and rarely sufficient. It helps against a claim that the buyer had no idea the system could be wrong. It does not neutralize a specific accuracy figure you published, because a buyer can reasonably read the general caution as coexisting with the specific number rather than cancelling it. Consistency between the two is what matters.
Our claims came from the model vendor's benchmarks. Are we covered?
Repeating someone else's number in your own marketing makes it your claim to your customer. Whether you can recover from the vendor is a separate contract question, usually governed by narrow indemnity and a cap at fees paid. Cite the source and the conditions if you use vendor benchmarks, and be clear you are describing the model rather than your product on the customer's data.
A customer says the AI got worse after an update. What is the exposure?
It depends on what you promised and how you handled the change. If you committed to specific behavior with no model-change rights and no notice, this is a straightforward breach discussion. If you described capability, reserved change rights, notified in advance and offered a testing window, it becomes a support matter. The drafting decides which conversation you are having.
Do we need different terms for enterprise customers?
You will end up with them regardless, because enterprise buyers negotiate. The risk to manage is drift: bespoke accuracy commitments granted deal by deal to close quarters, tracked nowhere, and inconsistent with the standard terms. Decide in advance which concessions are available and keep a register of what each customer was actually given.
Does a free tier or beta label change the analysis?
It helps and does not immunize. Absence of payment weakens contract-based claims, and a clearly labeled beta sets expectations that matter. But misrepresentation and unfair-practices theories are not limited to paying customers, and a free tier that funnels into a paid one is part of the same commercial pitch.
Where do most companies actually get caught?
The gap between the demo and the deployment. The demo used clean, curated inputs; production has scanned faxes, edge cases and a document type nobody tested. Every complaint has the same shape — it worked in the demo — and the fix is upstream: demo on realistic data and scope the claim to what you measured.
Every Claim on Your Site Is a Candidate Warranty
Accuracy numbers, capability tables, compliance badges and case-study figures live on pages most teams last reviewed at launch. They are the exhibits in any dispute about what you promised, and the first place a buyer's counsel looks.
See what your site currently claims. Run a free scan and review every page that makes a specific promise about your AI features.