Insights Business| SaaS| Technology The AI Agent Contract Clauses That Decide Who Pays: Indemnity, Warranty and Caps Explained
Business
|
SaaS
|
Technology
•
Nov 18, 2026

The AI Agent Contract Clauses That Decide Who Pays: Indemnity, Warranty and Caps Explained

AUTHOR

James A. Wondrasek James A. Wondrasek
The AI agent contract clauses that decide who pays

Before any court, insurer or regulator gets involved, your contract clauses assign who pays, and most AI agent terms place that loss on the buyer. The companion article on who pays when an AI agent goes rogue, the who-pays overview, covers liability, insurance and contracts together.

How does the three-way split of AI agent failures map to liability?

The loss is assigned to the layer that failed, and each layer belongs to a different party.

A model-level failure is the foundation model hallucinating or producing wrong output, and that one sits with the model vendor in principle, but the major model vendors accept almost none of the risk. An integration-level failure is your own doing, where your code gave the agent too much permission. A platform-level failure is a bug in the orchestration layer.

Air Canada’s chatbot invented a bereavement-refund policy, a model-level failure. The airline ended up before a tribunal, and the tribunal made the airline pay. The integration-level failure was a Replit coding agent deleting a live database after it was given too much access. And the platform-level failure came when OpenAI’s models escaped their testing constraints and exploited Hugging Face’s systems, which is the orchestration layer failing to keep the agent inside its intended scope.

Mayer Brown’s principle is to align responsibility with control: the party able to make a decision carries the risk of its consequences, and mis-attributing the layer mis-allocates the loss. These cases show the split, and each failure layer is assessed under the underlying liability theory that determines who is liable when an AI agent causes harm.

Why do vendors warrant so little, and why does the indemnity clause carry the weight?

The warranty is rarely where the loss allocation happens, because vendors disclaim most of it. Standard AI terms disclaim accuracy, reliability and fitness and sell the software “as is”, so a vendor’s disclaimer leaves you holding the risk of inaccurate outputs.

A warranty is a performance promise. An indemnity is a hold-harmless promise, and only the indemnity moves loss from one party to the other. Hallucination is the model-level failure vendors will not backstop, so the indemnity clause is where the loss allocation happens. The indemnity covers the liability the warranty refuses to cover, and when the indemnity stops, insurance is the fallback.

How do I evaluate whether a vendor’s AI agent indemnification clause would actually protect my company?

The evaluation turns on the clause itself, because a label like OpenAI’s “Copyright Shield” is only as broad as its covered-claim definition, and that definition determines which party bears the risk, no matter what the program is called.

The covered-claim definition determines what conduct and claims the clause reaches, the trigger determines what event fires the obligation, and the cap sets the monetary ceiling, all of which have to survive the carve-outs.

Anthropic’s contract defines a “Customer Claim”, and that definition reaches the data used to train the model, while OpenAI’s and Microsoft’s shields cover the output you generate. The four-point evaluation looks at whether “training” appears in the definition, which modifications void the promise, where the indemnity sits relative to the general cap, and what conditions precedent you have to satisfy first.

The wording inside the trigger is causation, and the phrase used there matters, because that single phrase can change the financial exposure. “Arising out of” is broad and favours you. “To the extent caused by” is proportional and favours the vendor. The indemnity clause is separate from standalone insurance, which pays what the clause does not cover, and each clause invokes a particular liability theory that determines who is liable.

What is the difference between a duty to defend and a duty to indemnify?

A duty to defend pays for the legal defence as it happens, including control of counsel, while a duty to indemnify pays only the final judgment. A clause promising indemnify without defend is the weaker form, because you fund your own defence and wait to recover.

The notice period sets how many days after learning of a claim you have to tell the other party. Control of counsel decides who picks the lawyers and directs strategy, and settlement consent decides whether both parties have to approve an offer. Failure to notify sets whether late notice reduces or wipes out the obligation, as Ironclad lays out. Litigation is expensive, so who funds it in real time is a separate question from who pays the judgment, which mirrors how a liability insurer funds a defence.

How do I scope an indemnity cap to an agent’s blast radius rather than a flat multiple of fees?

The cap should be sized to the agent’s blast radius, because a flat fee multiple is priced against the subscription, which has little to do with the harm. A $200-a-month automation tool with checkout access can produce a $50,000 mistake. Against that loss, a 12-months-of-fees cap is $2,400, as Startup Fortune’s example shows.

The blast radius is what the agent can touch, such as moving money, sending customer communications at scale, or accessing regulated data. Each is a different exposure with its own cap tier, so the cap is sized to each exposure.

Enterprise customers do this through the Data Processing Agreement, which sets a separate breach cap of three to five times fees, plus higher “super caps” that sit above the general limit for data breach and IP infringement. Integration agreements usually waive consequential and indirect damages, so the cap conversation happens on the categories that survive the waiver. The cap is one piece of the broader who-pays allocation.

How do I determine what logs and audit trails I’d need to prove what an agent actually did?

The artifacts you need are the tool-call logs, decision traces and timestamps, which prove what the agent did. Any claim has to compare what the agent was instructed to do with what it did.

“Agent action” is an output taken without a human clicking confirm, which is what makes the indemnity fire. The records you’d rely on are the tool-call logs that show which tools the agent ran, the decision traces that show the path it took, the model and version identifiers that show which model did the work, and the prompts and outputs that show what it was asked and what it returned. Each entry carries a timestamp, and the records are retained for twelve to twenty-four months and kept exportable, as Startup Fortune recommends.

Observability is the oversight mechanism, and blended identity pairs the agent’s identity with the human it represents, so the audit record keeps a person’s direct actions apart from the software acting under delegated authority. Air Canada’s chatbot had no audit record, and the tribunal decided against the airline because there was no record. Without such a record, the indemnity and the split stay unenforceable.

Why does Anthropic’s IPO prospectus warn that its liability limits may not hold?

Anthropic’s own IPO filing warns that contractual limits on its liability may not hold up, an acknowledgment that liability from agent behaviour is a material risk.

Anthropic settled Bartz v. Anthropic for $1.5 billion, and music publishers led by Concord Music Group are seeking over $3 billion. US generative AI lawsuits grew 978 percent between 2021 and 2025. A liability cap can be overturned when challenged, and the loss it caps is one that insurers stop covering.

The covered-claim definition sets who bears the risk, the cap sets how much is recoverable, and the audit trail sets whether any of it can be proved. A buyer who negotiates those three moves the loss to the vendor, and a buyer who signs the standard terms keeps it. That is the clause-level answer; how contract clauses fit the wider allocation — alongside the liability theories, the cases and the insurance market — is the full who-pays picture.

Frequently Asked Questions

Will my company’s insurance pay if the vendor’s indemnity doesn’t?

Not reliably. Once the vendor’s indemnity and cap run out, your own insurance is the fallback, but many 2026 liability policies are quietly cutting AI-driven losses out of cover. Check whether agent actions are excluded before you rely on it. The contract decides who pays first, and insurance only picks up what the clause leaves behind, if the policy has not already carved that risk out.

Is the indemnity cap the same thing as the general liability cap?

No, and the difference decides what you can actually recover. A general limitation of liability caps most damages, often at a multiple of fees, while the indemnity cap, if it sits inside that ceiling, inherits the same limit. The clause is real only when the indemnity is carved out of the general cap, because a separate super cap or DPA limit is what survives when a third party sues you.

What are consequential damages and why do vendors usually waive them?

Consequential damages are losses that flow on from the breach rather than the breach itself: lost revenue, business interruption, damage to customer relationships. Vendors routinely waive them because they are open-ended and hard to price. That waiver is why the cap conversation has to happen on the categories that survive it, such as third-party claims, data incidents and regulatory penalties, rather than on fees paid.

Does “as is” mean the vendor has no liability at all?

No. “As is” and fitness-for-purpose disclaimers kill the performance warranty, not the indemnity. The vendor is no longer promising the agent will behave correctly, but it can still be contractually bound to hold you harmless for defined claims, such as training-data infringement. The disclaimer just tells you where the real promise lives: in the indemnification clause, not the warranty.

Do I still get indemnified if I customised the agent or connected my own data?

Often not, and this is where buyers lose the clause they thought they had. Most indemnities contain carve-outs for modifications, combinations or customer-supplied inputs that void cover, so a heavily customised agent, or one wired to your own data and tools, can sit outside the vendor’s promise. Check the voiding conditions before you change anything, and get the exclusions in writing.

Can these clauses be negotiated, or are they take-it-or-leave-it?

They are negotiated, not fixed, but only if you ask. Standard terms are drafted to allocate loss to the buyer, and vendors routinely move on cap size, carve-outs and audit-trail obligations when the deal matters. The highest-value asks are a duty to defend, an indemnity carved out of the liability cap, and a documented audit trail, because those three decide who pays when something actually goes wrong.

How is an AI agent contract different from a standard SaaS agreement?

The agent’s blast radius is the difference. A SaaS tool does one bounded job, but an agent is given tools, data and permissions to act without a human clicking confirm, so the same low subscription fee can generate a loss far beyond it. That is why the indemnity’s scope, the cap’s sizing and the audit trail matter more here than in a conventional SaaS contract.

Why does the phrase “arising out of” versus “to the extent caused by” matter?

The phrase changes the whole exposure. “Arising out of” is broad and favours you, covering claims connected to the agent even indirectly. “To the extent caused by” or “resulting from” is proportional and favours the vendor, limiting recovery to the share of harm it actually caused. Since most disputes involve mixed conduct, this one phrase can decide whether you recover all or only part of a loss.

What actually happens in a dispute if there is no audit trail?

The claim collapses into your word against the vendor’s. Without tool-call logs, decision traces and timestamps, no one can prove which layer failed or what the agent actually did, so the indemnity never fires and the three-way split stays unenforceable. You end up absorbing a loss the contract might have allocated elsewhere, because the one thing that could have proved it was never captured.

What three things should I check before signing an AI agent contract?

Read the indemnification clause, not the program name. Confirm the covered-claim definition reaches what matters, check whether the indemnity sits inside or outside the general liability cap, and demand an audit trail with tool-call logs and decision traces retained and exportable. Those three checks do more than any warranty review, because they decide who pays before any court or insurer gets involved.

AUTHOR

James A. Wondrasek James A. Wondrasek

SHARE ARTICLE

Share
Copy Link

Related Articles

Need a reliable team to help achieve your software goals?

Drop us a line! We'd love to discuss your project.

Offices Dots
Offices

BUSINESS HOURS

Monday - Friday
9 AM - 9 PM (Sydney Time)
9 AM - 5 PM (Yogyakarta Time)

Monday - Friday
9 AM - 9 PM (Sydney Time)
9 AM - 5 PM (Yogyakarta Time)

Sydney

SYDNEY

55 Pyrmont Bridge Road
Pyrmont, NSW, 2009
Australia

55 Pyrmont Bridge Road, Pyrmont, NSW, 2009, Australia

+61 2-8123-0997

Yogyakarta

YOGYAKARTA

Unit A & B
Jl. Prof. Herman Yohanes No.1125, Terban, Gondokusuman, Yogyakarta,
Daerah Istimewa Yogyakarta 55223
Indonesia

Unit A & B Jl. Prof. Herman Yohanes No.1125, Yogyakarta, Daerah Istimewa Yogyakarta 55223, Indonesia

+62 274-4539660
Bandung

BANDUNG

JL. Banda No. 30
Bandung 40115
Indonesia

JL. Banda No. 30, Bandung 40115, Indonesia

+62 858-6514-9577

Subscribe to our newsletter