How a Liability Waterfall Actually Works Across Your AI Vendor Stack – Startup Fortune

When an AI system built from several vendors fails, the startup operating it will usually face the customer first. Contracts may assign responsibility among providers, but they rarely establish a coordinated payment order. That leaves the deploying company responsible for losses that vendor liability caps, exclusions and indemnities do not cover.

What Is a Liability Waterfall?

A liability waterfall is a contractual framework defining which party responds first when an AI system causes damage. It should connect responsibility to specific failure types, such as:

• Incorrect or outdated third-party data
• Faulty routing, tool selection or workflow logic
• Hallucinated or otherwise inaccurate model output
• Security failures or unauthorized access
• Actions taken without required human approval

Without this framework, each vendor relies on its own agreement. Those agreements often protect the vendor independently without addressing how its obligations interact with the rest of the stack.

Why the Customer-Facing Company Pays First

A 2024 Canadian ruling involving Air Canada illustrated the core risk. The airline’s chatbot provided incorrect information about bereavement fares, and a tribunal held the airline responsible for information presented through its website. The company could not avoid responsibility by treating the chatbot as a separate source.

The lesson is straightforward: customers generally pursue the business that sold the product, made the promise or deployed the agent. Whether that business can recover money from an AI vendor is a separate contractual question.

How Liability Usually Falls Across the Stack

A typical AI agent may combine a foundation model, an orchestration framework and one or more data providers. Each layer presents different recovery challenges.

Data Providers

A data provider may be responsible if its information was demonstrably inaccurate. However, many data agreements disclaim completeness, timeliness or accuracy. Recovery can therefore depend on whether the contract contains an enforceable data warranty and an adequate liability cap.

Orchestration Providers

The orchestration layer determines when the model runs, which tools it calls and what information reaches the user. Proving that a loss resulted from routing logic rather than model behavior, data quality or the startup’s configuration can be difficult.

Open-source software may also be supplied without warranties. Managed orchestration products commonly use standard software-as-a-service limitations tied to fees paid.

Foundation Model Providers

Foundation model providers commonly cap liability at an amount connected to recent service fees. For a startup with modest API spending, that cap may be far below the value of a customer claim or operational loss.

Providers may also exclude indirect, consequential or lost-profit damages. Even when the model contributed to the failure, the recoverable amount may cover only a small portion of the startup’s exposure.

The Deploying Startup

The startup is left with any amount that vendors do not pay. In practice, that may be most of the loss because the startup controls the customer relationship, system configuration, safeguards and final deployment decision.

How to Build a Contractual Waterfall

1. Assign Responsibility by Failure Type

Do not rely only on a general statement that each vendor is responsible for its services. Define categories of failure and identify which party bears primary responsibility for each one.

For example, a data provider could respond first to a breach of an express data-accuracy warranty. An orchestration provider could respond to failures in its routing logic, while a model provider could respond to qualifying generation failures attributable to its service.

2. Align the Separate Agreements

The vendors probably have no contracts with one another. The startup must therefore negotiate consistent language into each agreement. Definitions, claim procedures and exclusions should not leave the same risk unassigned—or allow every vendor to blame another layer.

3. Negotiate Meaningful Liability Caps

A cap based on fees paid may be acceptable during testing but inadequate once an agent handles financial transactions, medical information or customer commitments. Seek higher caps for high-risk failures, confidentiality breaches, security incidents and indemnified claims.

4. Request AI-Specific Indemnities

Intellectual-property indemnification alone does not cover harmful hallucinations, unauthorized actions or inaccurate advice. Ask whether indemnification applies to incorrect output, integration failures, privacy violations and third-party claims resulting from the vendor’s technology.

Also review exclusions and control-of-defense provisions. A broad indemnity has limited value if another clause removes the most likely AI failure modes.

5. Verify Insurance Coverage

Require appropriate vendors to provide evidence of technology errors-and-omissions, cyber or other relevant insurance. The startup should also carry coverage sized to its real downstream exposure rather than assuming vendor policies will protect it.

6. Preserve Technical Evidence

A waterfall cannot work without proof. Maintain model versions, prompts, retrieved data, tool calls, approval records and final outputs. These records help determine whether the failure originated in the data, orchestration, model or deployment controls.

The Practical Contract Review

Review every agreement in the stack and ask:

• What failures does this vendor accept responsibility for?
• What damages are excluded?
• How is liability capped?
• Does the indemnity cover likely AI incidents?
• Does the vendor maintain adequate insurance?
• How will responsibility be proven?
• Do the agreements leave any risk unassigned?

If the contracts do not connect responsibility across the stack, there is no true liability waterfall. There are only separate vendor protections—and a gap that the customer-facing startup will probably have to absorb.