The Warranty Problem: Who Pays When Your AI Feature Is Wrong?

Every software warranty ever written assumed deterministic behavior. You ship a function, it returns the same output for the same input, and your warranty covers the gap between documented behavior and actual behavior. AI features shatter that assumption entirely.

When your LLM-powered feature tells a customer something wrong — and that wrong thing costs them money — traditional warranty language leaves everyone pointing fingers at everyone else.

This is not hypothetical. Cumulative generative AI lawsuits in the U.S. passed 700 between 2020 and 2025, with year-over-year filings accelerating by 137%. The legal infrastructure governing software liability was built for a deterministic world, and the mismatch is already causing real damage.

The Air Canada Precedent: You Own Your Chatbot's Promises

The case that made every product team pay attention landed in February 2024. Jake Moffatt asked Air Canada's chatbot about bereavement fares. The chatbot said he could book a full-price ticket and apply for a bereavement discount retroactively. That was wrong — Air Canada's actual policy explicitly excluded retroactive applications.

When Moffatt requested his partial refund, Air Canada refused. Then they made an argument that should concern every company shipping AI features: they claimed the chatbot was a "separate legal entity" responsible for its own actions. The British Columbia Civil Resolution Tribunal rejected this completely — Air Canada is responsible for all information on its website, whether from a static page or a chatbot. The airline paid $812 in damages, a small amount that established an enormous principle.

If you deploy an AI system that makes claims to your customers, those claims are your claims. The model's non-deterministic nature is not a defense. The fact that you didn't write the specific words is not a defense. You chose to put the system in front of customers, and you own what it says.

Why "AI-Generated — Verify Independently" Doesn't Protect You

The most common risk mitigation strategy for AI features is a disclaimer: "This content was AI-generated. Please verify independently." In unregulated consumer contexts, this might reduce some exposure. In regulated industries, it is legally insufficient — and increasingly insufficient everywhere else too.

Consider the regulatory landscape:

• Healthcare: California's AB 3030 requires health care providers using generative AI for patient communications to include specific disclaimers and instructions for contacting a human provider. But the disclaimer is a floor, not a ceiling — only physicians, not AI systems, can make final decisions regarding medical necessity.
• Financial services: AI models used for credit scoring or risk assessment must comply with the Fair Credit Reporting Act. A disclaimer doesn't exempt you from substantive compliance obligations.
• Sales and marketing: The FTC requires that all claims — including AI-generated ones — be truthful and not misleading. Slapping a disclaimer on a product recommendation that your AI fabricated doesn't satisfy this requirement.

The pattern is clear: regulators treat disclaimers as a minimum disclosure obligation, not a liability shield. The substantive duty of care remains with the deployer. You cannot outsource professional judgment to a language model and then disclaim the consequences.

This matters practically because many teams treat the disclaimer as the end of their liability analysis. It should be the beginning.

The Insurance Gap Nobody Talks About

Here is a fact that should alarm any engineering leader shipping AI features: your company's existing insurance policies probably do not cover AI-specific liabilities.

Traditional errors and omissions (E&O) and general liability policies were written for deterministic software failures — bugs, outages, data breaches. They typically do not cover:

• AI hallucinations leading to financial losses
• Intellectual property infringement by generated content
• Defamation or misinformation from AI outputs
• Professional malpractice claims arising from AI-generated advice

This problem is getting worse, not better. In January 2026, Verisk released new general liability exclusion forms specifically for generative AI. Policies renewing in Q1 and Q2 of 2026 are the first to incorporate these exclusions. If your company renewed its general liability policy recently, check whether it now explicitly excludes AI-related claims.

The insurance industry calls this "silent AI" — AI exposures that are neither explicitly included nor excluded in existing policies. When a claim lands, both sides argue about whether coverage applies, and the policyholder usually loses.

Some specialized AI insurance products are emerging. Armilla, backed by Chaucer and Axis Capital, launched coverage in 2025 that requires ongoing model quality assessments. Testudo launched in January 2026 with a claims-made product for enterprises deploying generative AI. But these are early-stage products with limited track records, and they require the kind of model monitoring and evaluation infrastructure that most teams haven't built yet.

The practical takeaway: before you ship an AI feature that makes claims, recommendations, or decisions affecting your customers, ask your legal team whether your current insurance covers the failure modes. The answer will likely change your risk calculus.

Contract Language Is Still Catching Up

Most SaaS agreements were designed for deterministic software. The warranty section typically says the product will perform in "material conformance with its documentation." This language breaks down immediately for AI features because the outputs are probabilistic — the same input can produce different outputs, and neither the vendor nor the customer can fully predict what the system will say.

AI vendors have responded with aggressive disclaimers. The standard posture looks like this: "THE SERVICE IS PROVIDED AS-IS, WITH ALL FAULTS. Outputs are generated through machine learning processes and are not tested, verified, endorsed or guaranteed to be accurate, complete or current by Provider."

This creates a strange situation: the vendor sells you a product that makes claims, recommendations, or decisions — while simultaneously disclaiming any warranty that those outputs will be correct. The customer bears all the risk for outputs they cannot control.

The market is starting to push back. As AI gets embedded in core business processes, buyers increasingly refuse the warranty-free posture. We're seeing more deals include:

• Performance benchmarks: measurable accuracy thresholds the AI must maintain
• Outcome-based SLAs: service levels tied to output quality, not just uptime
• Remediation obligations: the vendor must fix or compensate when outputs fall short
• Audit rights: the customer can inspect model behavior and governance practices

For agentic AI — systems that take actions autonomously — the shift is even more dramatic. Mayer Brown's February 2026 analysis noted that agentic AI contracts are moving beyond traditional SaaS models entirely, incorporating elements of business process outsourcing (BPO) agreements: service definitions, broader indemnification, and governance frameworks that acknowledge the AI is acting, not just computing.

If you're integrating an AI vendor into your product, your contract should reflect who bears the risk when the AI is wrong. If the contract says "AS-IS" and you're putting those outputs in front of your customers, you've accepted all the liability with none of the control.

The Federal Response: AI LEAD Act

The regulatory gap hasn't gone unnoticed at the federal level. In September 2025, Senators Durbin and Hawley introduced Senate Bill 2937, the "AI LEAD Act," which would create a federal cause of action in product liability against developers and deployers of AI systems.

The proposed liability framework covers:

• Design defects: the AI system was unreasonably dangerous in its design
• Failure to warn: inadequate disclosure of known limitations or risks
• Breach of express warranty: the system didn't perform as represented
• Defects at deployment: unreasonably dangerous conditions present when shipped

At the state level, over 1,000 AI-related bills were introduced during the 2025 legislative session alone. States are authorizing private rights of action under AI-specific statutes, creating novel liability pathways that didn't exist two years ago.

The direction is clear: legislation is moving toward treating AI systems more like products than services, with the corresponding product liability obligations. The "it's just software" defense has a limited shelf life.

What Engineering Teams Should Do Now

The warranty problem isn't just a legal issue — it's an engineering issue. The decisions you make about monitoring, evaluation, and guardrails directly determine your company's liability exposure. Here's what matters:

Build evaluation infrastructure before you ship. If you can't measure your AI feature's accuracy on an ongoing basis, you can't set warranty thresholds, you can't prove compliance, and you can't defend against claims. This isn't optional instrumentation — it's the foundation of your legal defense.

Classify your outputs by liability risk. An AI that suggests blog post titles has a different risk profile than one that provides medical billing codes. Your guardrails, monitoring, and human-in-the-loop requirements should scale with the cost of being wrong.

Treat AI outputs as untrusted input at system boundaries. The same way you wouldn't trust user input in a web form, don't trust model outputs when they flow into customer-facing surfaces, financial calculations, or contractual commitments. Validate, constrain, and audit.

Work with your legal team on contract language now. If you're buying AI from vendors, push for performance warranties. If you're selling AI features, understand what you're implicitly warranting by putting outputs in front of customers. The "AS-IS" disclaimer may not survive the next wave of legislation.

Check your insurance. This is a five-minute conversation with your broker that could save your company millions. Ask specifically about AI exclusions in your current policies and whether you need specialized coverage.

The warranty problem for AI features is the gap between what customers reasonably expect and what the legal framework currently guarantees. That gap is closing — through litigation, legislation, and market pressure. The teams that build for accountability now will have a significant advantage over those who wait for a lawsuit to force the issue.