Guides

    The four-clause AI addendum for client work

    Disclosure, IP warranty carve-out, training-data exclusion, liability cap. What each clause stops, the order to negotiate them, and where the addendum belongs.

    Versely Team8 min read

    The useful thing about contract advice for AI production work is how quickly it converged. Published guidance aimed at creative studios keeps landing on the same short list of what an existing services agreement needs bolted onto it. Agent Mode AI's operator write-up sets the minimum at four: disclosure of AI use, an IP warranty carve-out, a training-data exclusion for client materials, and a liability cap tied to the fee paid. Numonic's version adds metadata preservation requirements and audit rights, which is really the same four plus the machinery that makes two of them enforceable.

    That convergence matters because it means you are not designing anything. You are picking a known set, understanding what each item stops, and negotiating them in an order that does not blow up the deal. This is a walkthrough of that, written from the supplier side. It is not legal advice, and an addendum going onto a contract with real exposure should be read by someone qualified in the relevant jurisdiction.

    What each clause actually stops

    Every one of the four exists because a specific, recurring failure happened to somebody. Reading them as abstract risk management makes them feel optional. Reading them as scar tissue makes them obvious.

    Clause The failure it prevents Who resists it
    Disclosure of AI use Client learns from a third party. The conversation stops being about method and becomes about what else you withheld Almost nobody, once framed correctly
    IP warranty carve-out You warranted originality and full assignment on output that may carry no copyright to assign Client procurement, hard
    Training-data exclusion Client brand assets end up in a model training set, breaching a duty they owe upstream Nobody. Clients usually ask first
    Liability cap tied to fee A modest deliverable generates exposure many multiples of the invoice behind it Client legal, moderately

    The disclosure clause is the cheapest and the one most often skipped, because omitting it feels safe in the moment. It is not, and the reason has nothing to do with ethics. The account-losing event is almost never a client discovering that you use AI. It is a client discovering it somewhere else, mid-campaign, with a question they cannot answer. There is a whole conversation to have about how you frame that on the kickoff call, and the clause is what makes the framing binding rather than a remembered chat.

    Note also that client-facing disclosure and audience-facing disclosure are different obligations with different enforcers. The first is commercial and entirely within the contract. The second is regulatory: Article 50 of the EU AI Act became binding on 2 August 2026 and turns on whether realistic synthetic content would falsely appear authentic to its audience. Your addendum should say which of you is responsible for the second one, because it is not automatically you.

    Sequence them in the order the client can say yes

    The order matters more than most people expect, because two of the four are easy and two are not, and leading with a hard one turns a twenty-minute review into a legal escalation.

    First, disclosure. It costs the client nothing, it makes you look like the adult in the room, and it establishes a fact you need for the third clause to be honest. You cannot carve out AI-generated portions of a deliverable in a contract that never admits there are any.

    Second, the training-data exclusion. This one usually arrives from their side anyway. Enterprise clients increasingly carry an upstream obligation that brand assets never enter a training set, and that duty flows down to you. Agreeing to it is free, and volunteering it before they ask is a small credibility win. The work is in honouring it, which is a stack audit rather than a drafting exercise.

    Third, the IP warranty carve-out. This is the hard one, and it is hard because it directly contradicts a sentence that has sat unexamined in services agreements for twenty years: the deliverables are original works, and all rights in them are assigned. Against generative output that sentence promises something you may not be able to deliver, because copyright in AI-generated material turns on the degree of human authorship and may simply not be there to assign. Settle the carve-out wording in advance with counsel rather than drafting it live on a review call.

    Which means a blanket originality warranty on AI-assisted deliverables is a promise you may already have broken. Leading with that framing rather than with proposed redlines is what gets it through. You are not asking for a concession, you are removing a warranty that was never accurate.

    Fourth, the liability cap. Leave it last and negotiate it with the commercial terms, not the technical ones, because it belongs to the same conversation as the fee. Capping at fees paid is the standard shape for a reason: a generated asset distributed across paid social can create exposure orders of magnitude larger than the invoice behind it, and no supplier prices for unlimited downside. Expect fair carve-outs, and expect the negotiation to be about which ones.

    The two clauses that turn four into six

    The extended set adds metadata preservation and audit rights, and both have become considerably less theoretical.

    Provenance metadata is now infrastructure rather than a nice-to-have: C2PA passed 6,000 members and affiliates as of January 2026, and the major creative and productivity suites now attach Content Credentials by default. That changes the handoff. If you deliver a file with credentials intact and the client re-encodes it through a compression step that strips them, they have created their own compliance problem, and the contract should say so before it happens rather than after. Understanding what survives a real pipeline and what does not is what turns the clause into a deliverable spec instead of a wish.

    Audit rights are the enforcement half of the training-data exclusion. A promise that client material never enters a training set is unverifiable unless someone can check, and the honest version of the clause names your tool list and commits to notice on change rather than promising universal compliance you cannot see. Clients who treat licensed training data as a buying criterion will ask for this. Give them a version you can actually satisfy.

    Where the addendum belongs

    Put it in the master agreement, not the statement of work, and reference it from every SOW. Three reasons.

    An SOW-level addendum has to be renegotiated per project, and it will eventually be omitted on a rushed one, which is precisely the project where it matters. A master-level addendum survives staff turnover on both sides. And the liability cap behaves differently depending on which document it sits in, so putting it where the commercial terms live keeps that interaction visible instead of accidental.

    Two drafting habits worth adopting. Define "AI-Generated Portions" once, early, and use the defined term everywhere, because vagueness in that definition undoes the IP carve-out. And make the disclosure clause reciprocal: you disclose what you used, and the client discloses any downstream regulatory or platform obligations they know about that the deliverable has to satisfy. That second half costs them nothing and has saved more campaigns than the first half.

    If your existing agreement already handles usage rights and licensing scope properly, the addendum is genuinely short. If it does not, fix the underlying licensing language first. An AI addendum bolted onto a contract with a vague rights grant is a good clause protecting a bad one.

    FAQ

    Do I need this for a small one-off project?

    The disclosure clause and the liability cap, yes, and both fit in a paragraph each. The IP carve-out and the training-data exclusion matter most where the deliverable is a durable brand asset or where the client has enterprise obligations of their own. A single social cutdown for a local business is a different risk profile from a brand mascot that will appear on packaging for three years.

    Will asking for a warranty carve-out cost me the deal?

    Rarely, if you lead with the reason rather than the redline. The framing that works: the existing warranty says something about copyright that current law does not support, and signing it unchanged puts both parties in a worse position than being accurate. Clients whose legal teams have looked at this already know. Clients whose legal teams have not will usually be relieved someone raised it.

    Does an addendum help if the client never tells their own client?

    Partly. It documents what you disclosed and when, which is the thing that protects you when the chain breaks further downstream. It does not make you responsible for their onward disclosure, and it should say so explicitly. Pairing the addendum with a written internal disclosure policy on their side is the version that actually closes the gap.

    Is four clauses enough for regulated sectors?

    No. Finance, health, legal services and anything touching children's advertising carry sector rules that sit on top of all of this, including substantiation requirements for claims and pre-clearance processes. The four-clause set is a floor for general commercial work. Treat the broader licensing and compliance picture as the starting point for anything regulated, and get counsel involved early rather than at signature.