Capping liability at the fee you were paid

Two facts worth holding together. A widely circulated Kalshi commercial was reported to have been produced for around $2,000 in total, a figure that would not have covered catering on a conventional shoot of the same reach. Coca-Cola's AI-assisted holiday campaign compressed a production timeline from about a year to about a month, then drew sustained public backlash large enough to become the story about the campaign.

Production spend has collapsed. Downside has not. A deliverable that takes four figures to make can, if it goes wrong in public, generate a remediation bill, a pulled media buy and a brand problem that no invoice in the relationship comes close to covering. That asymmetry is why a liability cap tied to the fee paid sits in the minimum viable clause set alongside disclosure, the IP warranty carve-out and the training-data exclusion.

This is a supplier-side guide to negotiating it. Not legal advice, and caps interact with mandatory local law in ways that need a qualified read.

What the cap does, and the four things it does not

A cap limits the money a client can recover from you for claims arising under the agreement. That is all it does, and being clear about the limits of it makes the negotiation easier, because clients often resist a cap for reasons the cap was never going to solve.

It does not stop a claim being brought. It does not prevent injunctive relief, so a client can still stop you doing something regardless of the number. It does not cap reputational consequences, which for most studios are the larger risk anyway. And in most jurisdictions it cannot cap certain heads of liability at all, however the contract is drafted.

There is also a quieter clause sitting next to the cap that frequently matters more than the cap itself: the exclusion of indirect and consequential loss. A mutual exclusion of lost profits, lost revenue, lost data and loss of goodwill removes the categories where the numbers get genuinely unbounded. A studio that wins a generous cap but signs up to consequential loss has negotiated the wrong clause. If you can only get one, take the exclusion.

Cap shapes and which one to open with

The fee-linked cap is standard because it is the only reference point both sides already agree on. Everything else is arbitrary. Within that, four shapes recur.

Shape	Where it fits	What to watch
Fees paid under the relevant SOW	One-off project work	Can be very small on a first engagement, which reads as unserious to procurement
Fees paid in the 12 months preceding the claim	Retainers and ongoing work	Grows with the relationship. Usually the workable compromise
A multiple of fees paid, commonly 1.5x or 2x	When the client will not accept 1x	Concede the multiplier long before conceding the principle
Greater of fees paid or a fixed floor	Small SOWs under a large master agreement	The floor is the client's protection against a token cap

Open at fees paid under the SOW if the work is project-based, and at trailing twelve months if it is a retainer. The trailing-twelve shape is the one that survives review most often, because it scales with exposure in a way both sides find intuitive: a client three years into a substantial relationship has more at stake and a correspondingly larger cap, without anyone renegotiating.

If you price per deliverable rather than per retainer, watch the interaction between the cap and the SOW structure. A cap written per-SOW across twelve small SOWs behaves very differently from one written across the master agreement, and a client who splits work into many small statements has, perhaps unintentionally, reduced your exposure to almost nothing. Expect that to be spotted and to trigger the floor.

The single most important operational rule: do not agree a cap larger than your professional indemnity cover. A cap is a promise to pay. If your insurance stops at a number and the contract does not, the gap is yours personally. Check the cover before the call, not after.

The carve-outs clients fairly demand

No cap is unlimited in scope. Some exclusions are genuinely reasonable, some are mandatory as a matter of law regardless of drafting, and a few are the client trying to reopen the whole negotiation through a side door. Knowing which is which is most of the skill.

Reasonable and usually non-negotiable:

Fraud and wilful misconduct. Frequently unexcludable anyway.
Death or personal injury caused by negligence. Unexcludable in many jurisdictions.
Breach of confidentiality. A confidentiality breach can be catastrophic and has nothing to do with the size of the fee.
Breach of the training-data exclusion. If you promised client material would not enter a training set, a cap at the project fee makes the promise close to worthless. Expect this one to be carved out or given its own higher sub-cap, and treat that as fair.
Personal data breaches. Data protection regimes carry their own liability and the contract cannot wish it away.

Reasonable to resist, or to convert into a sub-cap rather than accept uncapped:

The IP infringement indemnity. More on this below, because it is the one that matters most in an AI addendum.
Any breach of the agreement. This is not a carve-out, it is the removal of the cap. If the redline says liability is capped except for breach of contract, the cap does nothing.
Regulatory fines incurred by the client. Fines usually follow the party that published. Where the labelling duty sits should be settled in the disclosure and indemnity clauses, not smuggled into the cap. Article 50 of the EU AI Act has been binding since August 2026 and lands primarily on whoever deploys the content.

The consistency trap worth avoiding

Here is the drafting error that undoes an otherwise well-built addendum.

You carve out the IP warranty because you cannot honestly warrant that AI-generated portions attract copyright, and because assigning rights that may not exist is a promise you cannot keep. Good. Then, four paragraphs later, you accept an uncapped IP indemnity covering all claims that the deliverables infringe or fail to convey the promised rights.

That indemnity re-imports the exact risk the carve-out removed, and removes the cap while doing it. The two clauses have to be read together or they contradict each other in the client's favour.

The coherent version: indemnify against third-party infringement claims arising from your inputs and your production choices, which is risk you controlled. Do not indemnify against the legal status of AI-generated material under copyright law, which nobody controls. If the client needs comfort on the second, offer a sub-cap at a multiple of fees rather than an uncapped promise, and be explicit that the sub-cap replaces rather than supplements the general cap.

That negotiation goes better if you have already shown your work. A documented quality-control process and a brand-safety checklist reframe the conversation from "how much will you pay when it goes wrong" to "here is why it does not." Procurement teams price process. Turning up with one is worth more than a paragraph of argument.

What to concede

You will not win every point, and holding the wrong ones costs deals. A rough priority order, most important first:

The consequential loss exclusion, mutual. Hold this hardest. It removes the unbounded categories.
The cap existing at all, tied to fees. Hold. An uncapped agreement is uninsurable in practice.
A cap no larger than your insurance cover. Hold. This is not a negotiating position, it is arithmetic.
The multiplier. Concede. 1x to 1.5x to 2x is a cheap ladder and it usually closes the point.
A floor on small SOWs. Concede, within reason. It is a fair ask and it signals confidence.
Sub-caps for confidentiality, data and the training-data exclusion. Concede, at a level you can insure.
Uncapped anything else. Decline, and explain the insurance reason rather than the legal one. It lands better and it is true.

Two habits make all of this easier. Quote the risk into the price rather than arguing about it abstractly: a base-plus-change-order structure gives you somewhere to put additional exposure when a client insists on a larger cap. And keep the addendum's clauses reading as one document, so the cap, the warranty carve-out and the indemnity split are consistent with each other. The broader licensing and compliance picture is worth having in view while you draft, because the cap is the last line of defence rather than the first.

FAQ

Is capping at the project fee too aggressive for a first engagement?

It is the correct principle but often an awkward number, because a first project can be small enough that the cap looks like a rounding error to a large client. That is what the floor is for. Offer the greater of fees paid or a fixed amount you can insure, and the objection usually disappears without you giving up the fee linkage.

Does the cap protect me if the client's audience reacts badly to AI in the creative?

Against damages claims under the agreement, within the cap and subject to carve-outs. Not against the commercial reality, which is the bigger risk. The perception gap is real and moving in the wrong direction: eMarketer reported a 37-point gap between advertising executives and Gen Z and millennial consumers on AI in ads, widened from 32 points in 2024. Handle that in the creative and in the disclosure approach, not in the liability clause.

Should the cap be mutual?

Ideally yes, and asking for it is a useful negotiating move even when you expect to lose it. Client-side liability under a production agreement is mostly the payment obligation, which is normally carved out of any cap anyway, so a mutual cap costs them little. Where it earns its keep is signalling that the clause is a standard commercial allocation rather than a supplier trying to escape responsibility.